All AI models

Multimodal mixture-of-experts model with a causal encoder-decoder design and separate computation budgets for reading input and generating text.

A native multimodal GLM model with sparse and linear attention.

A lower-cost model for focused tasks at high volume.

A hosted general-purpose model for text, image understanding, tool calls and long tasks.

A multimodal mixture-of-experts model with 600B total and 27B active parameters for text, image and video tasks.

A model for extended coding and knowledge work with vision and computer-use support.

A multimodal coding and agent model with sparse attention, quantization-aware training and a DSpark draft head.

Hy4 Preview is a text model with a 770B mixture-of-experts backbone and 49B active parameters per token.

GPT-5.6 Luna targets repeated tasks with low token costs.

A model for coding, tool use and complex professional tasks.

A 120B-total, 12B-active vision-language model that combines a hybrid language backbone with a C-RADIO encoder.

A text model with stronger post-training on the GLM-5.2 base.

A K3-derived multimodal model with sparse attention, 2.8T total parameters and 104B active parameters.

A hybrid LatentMoE reasoning model with 550B total and 55B active parameters.

Hy3 is a 295B mixture-of-experts language model with 21B active parameters and a separate 3.8B multi-token prediction layer.

A multimodal text-output model with a 1,048,576-token context and selectable reasoning effort.

A model for high-volume classification, extraction and routing tasks.

A multimodal agent model that accepts text, images, video and audio and returns text.

A model for coding, writing and agent tasks with configurable reasoning.

A long-context Kimi model with always-on reasoning and a million-token API context.

Multimodal Flash model with adjustable thinking, computer use, code execution and search tools.

Large text reasoning model with sparse attention, adjustable thinking effort and a separate general-availability release for agent tasks.

Multimodal text-generation model with text, image, video, audio and PDF input and a shared 1,048,576-token window.

A hybrid text model with Mamba-2, expert and attention layers, 30B total parameters and 3B active parameters.

MiMo V2.6 Pro accepts text, images, video and audio, then returns text.

A model for visual reasoning, screen interaction and writing with optional reasoning.

A long-context GLM release with a shared sparse-attention indexer.

Hosted Flash-Next model with built-in tools and a default one-million-token context.

Dense 27-billion-parameter vision-language model with weights released under Apache 2.0.

A native multimodal model with sparse attention that accepts text, images and video.

Flash preview model with multimodal input, text output and a 1,048,576-token input limit.

Claude Sonnet 5 is an active earlier Sonnet release with adaptive thinking.

Multimodal model for translation, transcription, extraction and routing with optional thinking.

Dense vision-language model with 30.7 billion parameters and a 256K context.

A 27B decision model adapted from Qwen3.8-27B.

Vision-language mixture-of-experts model with 25.2 billion total parameters and 3.8 billion active parameters.

A visual reasoning model with a 196B language backbone and 1.8B vision encoder for images, video, code and tool calls.

A reasoning model for complex problems and extended coding tasks.

gpt-oss-120b is a downloadable text reasoning model with 117 billion total parameters and about 5.1 billion active parameters.

GPT-5.6 Sol is the main professional-work model in the GPT-5.6 family.

Multimodal text-generation model retained for existing active users.

A 12B multilingual text model developed jointly with NVIDIA, with pretrained and instruction checkpoints.

A reasoning model for complex tasks and work that spans many steps.

A sparse vision-language release with preserved reasoning context.

Low-latency text-generation model for high-volume processing of text, images, video, audio and PDFs.

GPT-6 Sol is a coding and agent model that remains available alongside GPT-6.1 Sol.

Earlier multimodal Flash text-generation model restricted to existing users during its deprecation period.

A text-and-image reasoning model with tool calling and structured output.

A July 2026 Opus model for reasoning, coding and visual understanding.

GPT-4o Mini is a deprecated small model for focused text and image tasks.

Claude Sonnet 4.6 remains active with a one-million-token context and adaptive thinking.

Eight-billion-parameter multilingual embedding model with task instructions and a 32,000-token context.

A large open-weight model that reads text, images and audio.

Coding model with 30 billion total parameters and three billion active parameters for repository and terminal tasks.

An earlier Opus model for coding, document processing and agent workflows.

A mid-tier GPT-5.6 model for visual understanding with configurable reasoning.

Hosted multimodal model with text, image and video input and a one-million-token context.

text-embedding-3-small provides text vectors for search and relatedness tasks.

A downloadable text reasoning model with 21 billion total parameters and about 3.6 billion active parameters.

A deprecated Sonnet model for reasoning, coding and visual understanding.

An earlier Haiku model with manual reasoning budgets.

A sparse multimodal model with a one-million-token context and 52B active parameters, including embeddings and the output layer.

A smaller multimodal Inkling model with public weights.

A text, image and video model with selectable thinking and non-thinking modes.

A multimodal text-output model with 31B total parameters that accepts video, audio, images and text.

A nine-billion-parameter vision-language release.

A multimodal mixture-of-experts model with 816B total and 42B active parameters that returns text.

Multimodal text model with reasoning, structured output, function calls and code execution.

Claude Opus 4.6 expanded the Opus context window and introduced adaptive reasoning in this release line.

Multimodal reasoning preview with a million-token input allowance and tool-based workflows.

GPT-5 Mini is a deprecated smaller reasoning release for defined tasks.

A deprecated compact GPT-4.1 model for instruction following and tool calls without a separate reasoning stage.

Grok 4.5 supports text and image input, reasoning and tool use.

GPT-5.4 nano is a deprecated model for small, repeated tasks such as classification, extraction and ranking.

GPT-5.4 Mini targets frequent coding, computer interaction and delegated tasks.

An open-weight text model for coding, reasoning and extended tool workflows.

A diffusion language model for interactive reasoning and tool workflows.

Vision-language model for coding, visual understanding and GUI tasks.

An open-weight coding model with a one-million-token context.

Smaller instruction-following reranker with long-document support.

GPT-5.4 is an available coding and professional-work model with a March 2026 snapshot.

A hosted Ultra decision model that accepts text and images through Liquid's decisions API.

Grok 4.6 is a text-and-image reasoning API with a 500,000-token context and configurable reasoning effort.

GPT-4.1 is a deprecated non-reasoning model for instructions, tools and long inputs.

Multimodal model that returns text from audio, video, image and text inputs, with 113 audio languages and dialects.

Grok 4.3 combines a one-million-token context with text and image input.

The largest core Gemma 3 vision-language model, with 27 billion parameters and text output from images and text.

A coding-focused multimodal model with required thinking mode and downloadable weights.

The smallest Llama 3.1 instruction model, with 8B parameters.

A sparse text model for reasoning and software engineering.

Text-only mixture-of-experts model with 2.4 trillion total parameters and 95 billion active parameters.

GPT-5.2 is an earlier professional-work release that remains available in the catalog.

The sparse published checkpoint corresponding to Qwen3.5 Flash.

A decision model that answers typed questions about shared input.

GPT-5.5 supports demanding coding and professional tasks through text and image input.

GPT-5 nano is a deprecated small model aimed at summarization and classification.

GPT-5.1 is a deprecated coding and agent model with optional reasoning.

A published coding-agent model with a hybrid architecture.

A GLM release trained for sustained multi-step work.

A reasoning and coding release with preserved thinking across turns.

Earlier multimodal Pro model for reasoning and coding, with access limited to existing users during its deprecation period.

Reranker that follows natural-language ranking criteria.

A 70B text-only instruction model for multilingual dialogue.

A 24B instruction-model revision that improves precise instructions, repetition control and tool-call formatting.

Claude Fable 5 is the earlier Fable release for extended reasoning.

A compact Qwen3.5-4B decision model trained using probability distributions from larger Decider teachers.

text-embedding-3-large converts text into vectors for matching, retrieval and grouping.

A multimodal model combining the V2 Flash language backbone with vision and audio encoders.

Multimodal K2 successor that combines image and video understanding with coding and reasoning.

A Muse Spark model that accepts text, images, video, audio and PDF through the API and Muse Code.

A hosted decision model that accepts text and images through Liquid's decisions API.

Claude Opus 4.7 is an active coding and reasoning model with higher-resolution image processing than Opus 4.6.

Smaller instruction-following reranker for interactive retrieval.

A 30B dense model distilled from Muse Spark that accepts text and images and returns text.

The large published vision-language checkpoint behind hosted Qwen3.5 Plus.

An earlier Mercury reasoning model with legacy API access.

A 16-expert Llama 4 model with 109B total and 17B active parameters.

Multimodal embedding model that maps text, images, video, audio and PDFs into one vector space.

Instruction-following reranker with a 32,000-token pair limit.

A 128-expert Llama 4 model with 400B total and 17B active parameters.

A model for software agents and editable office work, including multi-agent teams and dynamic tool search.

A smaller instruction-tuned checkpoint for coding agents.

Text embedding model with a 2,048-token input limit and output dimensions from 128 to 3,072.

A 0.6B-class embedding model for independent text queries and documents.

Text-only model for summaries, translation and classification.

V2.5 Pro is a text-only mixture-of-experts release for software engineering and long agent runs.

GPT-4o is a deprecated general model with text and image input and text output.

A text reasoning model with 120B total and 12B active parameters.

An Apache-licensed multimodal model that combines instruction following, reasoning and coding with selectable reasoning modes.

A dense vision-language release with preserved thinking context.

GPT-5.3-Codex is a deprecated model specialized for coding agents.

A 24B text model that moves the Small line to Apache-licensed base and instruction weights.

A GLM text release with a 200,000-token context.

Twelve-billion-parameter Gemma 3 vision-language checkpoint with instruction tuning and a 128K context.

Four-billion-parameter text embedding model for instruction-aware retrieval across more than 100 languages.

Updated full-size R1 reasoning model with more post-training computation, deeper reasoning and revised prompt handling.

A smaller mixture-of-experts GLM release for local deployment.

Claude Opus 4.5 is an active earlier Opus model with manual extended thinking.

A coding and office model trained with reinforcement learning across more than 200,000 real-world coding environments.

An August 2025 multimodal Medium update that changes tone and performance while retaining 128K context.

A text model with agent-focused post-training and a 131,072-token context.

Eight-billion-parameter multilingual reranker for instruction-guided relevance scoring.

A text embedding model that produces fixed 1,024-dimensional vectors for search and classification.

General text embeddings compatible with the other Voyage 4 models.

GPT-5 is a deprecated general reasoning and coding release.

A 4B-class embedding model for independent text queries and documents.

An 8B dense text model derived from Granite 4.1 Base with full, low-effort and direct-answer thinking modes.

Nano Banana 2 image model for generation and editing from text, images, video and PDFs.

A smaller Laguna release for coding agents on local hardware.

Smaller hosted text encoder in the shared Voyage 4 family.

A mid-sized sparse release in the Qwen3.5 vision-language family.

The non-thinking instruction release of Qwen3 Next.

A text model with hybrid Mamba-2, expert and attention layers and 3.5B active parameters.

Multimodal model that combines native audio understanding with image, video and text input.

A 123B coding-agent model for tool use and changes across repository files, with weights under a Modified MIT license.

A dense Qwen3.5 release with image understanding.

Lower-latency image model with reference generation, marked-up editing and layer decomposition.

Mixture-of-experts text embeddings with a shared Voyage 4 vector space.

The smallest Ministral 3 release pairs a 3.4B language model with a 0.4B vision encoder and offers multiple training variants.

Four-billion-parameter vision-language model with image and text input and a shared 128K window.

A smaller sparse model in the GLM-4.5 family.

A downloadable 20B-class gpt-oss adaptation for policy-based text classification.

GPT Image 2.5 Sunburst is an image generation and editing model aimed at precise visual changes.

Image generation and conversational editing model with multimodal references and character consistency across editing turns.

Seed 2.1 model for agent workflows and multimodal understanding.

Compact multimodal model designed for high request volumes, batch generation and dense deployment.

A 24B reasoning model with 2B active parameters.

A search-grounded model available through Perplexity's Agent API.

Grok Build 0.1 is the coding-focused API model behind the current Grok Code Fast aliases.

The largest Ministral 3 model provides 14B text-and-image checkpoints for base training, instruction following and reasoning.

An 8B multimodal Ministral release with separate pretrained, instruction and reasoning checkpoints for local deployment.

An earlier Muse Spark API model with a 1,048,576-token context and Standard data mode.

Revised non-thinking K2 model with a doubled context window and a stronger focus on agentic coding.

Multimodal model for document processing and automation with text, image and video inputs.

A dense 128B multimodal model that combines instruction following, reasoning and coding with optional reasoning per request.

A hybrid reasoning model with sparse experts.

Quality-focused search reranker with Context Focus for returning relevant document passages.

A 30B-A3B translation model for 33 languages that accepts instructions written in multiple languages.

GPT Image 2.5 Flare targets everyday image creation with short iteration cycles.

Low-latency search reranker with calibrated relevance scoring and Context Focus.

Compact Nano Banana image model optimized for 1K output and simple conversational edits.

Sparse multimodal model with 218 billion total parameters, 25 billion active parameters, tool use and 48 languages.

An image-generation and editing model with reference-image input and multi-turn editing.

text-embedding-ada-002 is an older text embedding model that remains in the public catalog.

An Apache-licensed multimodal expert model with 675B total parameters and 41B active parameters per token.

Distilled 5.6-billion-parameter image transformer for generation, editing and compositions with up to ten references.

GPT Image 2 creates and edits images with flexible dimensions and detailed reference-image input.

A 14B text model with a 16K context that provides the base for later Phi-4 reasoning releases.

Nano Banana Pro model that generates text and images from text and image inputs.

An 8.38B relevance model that scores text queries against text, image or combined documents.

Code embeddings trained with completed pull requests and repository tasks.

Step 3.5 Flash is a text reasoning model for code and tool use.

An earlier Mercury reasoning model with legacy API access.

The middle-sized Llama 3.1 instruction model, with 70B parameters.

Long-reasoning K2 model trained to alternate internal reasoning with repeated tool interactions.

Image model for multi-image editing, consistent references and dense text rendering.

A research-oriented Grok service that uses multiple agents in parallel.

A Llama 3.1 405B derivative with optional reasoning and structured output.

Shared embeddings for text, images and video frames.

Image generation and editing model for text-rich graphics, marked-up edits and layer separation.

Image model with visual reasoning and search-enhanced generation.

First-generation multimodal Nova model that accepts text, images and video.

The larger Llama 3.2 text assistant, with 3.21B parameters.

o3 is a deprecated reasoning model for multi-step work across text, code and images.

A multimodal model built from GLM-4.5-Air.

Speech synthesis model for 130 languages with expressive delivery, speaker consistency and custom voice controls.

A coding model optimized for frequent editor completions and fill-in-the-middle requests.

Low-latency embedding model that shares its vector space with Embed 5 Pro.

A 4B Gemma-derived classifier for multimodal safety assessment and custom-policy reasoning.

Four-billion-parameter image generation and reference-editing model distilled to four sampling steps.

High-throughput speech model for single-speaker reading in 101 languages.

A code-retrieval embedding model with selectable vector dimensions and precision.

An image-generation model that accepts text prompts and reference images.

UI-TARS 1.5 is a downloadable vision-language agent for screen-based tasks.

A downloadable model for multilingual software work, code review, tool use and longer plans.

A multimodal enterprise model offered through hosted APIs and negotiated private deployments.

GPT-4o Mini Transcribe is the smaller GPT-4o-based transcription release.

GPT-Audio Mini is a deprecated lower-cost audio model for Chat Completions.

Hosted image generation and editing model with a fixed-snapshot standard endpoint.

Second-generation Titan text-vector model available through Amazon Bedrock.

Quality-focused embedding model for text, images and mixed documents, sharing a vector space with Embed 5 Fast.

An open-weight Hermes assistant derived from Meta's Llama models.

Dedicated programming model in the Seed 2.0 family.

GPT Image 2.6 Sunburst creates and edits images for detailed visual compositions.

A coding and tool-use model with 230B total parameters and 10B active parameters per token.

A 24B Small update that adds image understanding and expands the context window to 128K.

A compact Llama 3.2 text assistant with 1.23B parameters.

Image generation and editing model distilled to eight denoising steps.

Compact Command R model for text generation, retrieval and tool workflows.

An Apache-licensed 24B audio-understanding model that answers questions about recordings and requests tools from spoken instructions.

Recorded-speech recognition model for more than 85 languages with speaker labels, word timestamps and custom vocabulary.

An open-weight Hermes assistant derived from Meta's Llama models.

A 123B text model with 128K context and public weights under the Mistral Research License.

Image generation and editing model for complex layouts and Chinese and English text.

August 2024 Command R text model for generation and retrieval workflows.

Image generation and editing model with up to ten reference images and bounding-box layout controls.

A 7B model that reads images, video and text and returns text.

A 1.55B embedding model that maps text queries and text or image documents into retrieval vectors.

First-generation multimodal Nova model designed to balance capability and serving cost.

GPT Image 2.6 Flare supports rapid image generation and editing.

A 111-billion-parameter text model for business workflows in 23 languages.

A sparse expert model with 141B total parameters, 39B active parameters and a 64K context.

GPT-Audio is a deprecated audio model accessed through Chat Completions.

GPT Image 1 Mini is the lower-cost image release from the GPT Image 1 generation.

GPT-5.5 Pro spends more computation on difficult reasoning tasks.

An instruction-tuned mixture-of-experts language model with 80B total and 13B active parameters.

A public-preview image model designed for faster, high-volume generation and editing.

An image generation and editing model in public preview through Microsoft Foundry.

GPT-4o Transcribe is a deprecated speech-recognition model based on GPT-4o.

A multilingual retrieval model that creates one vector per text.

August 2024 Command R+ text model for retrieval and multi-step tool workflows.

A 21B text reasoning model trained from scratch.

High-capacity hosted image model for generation and reference-image editing.

A preview speech model that combines transcription and diarization with speech translation and audio question answering.

A 1.8B translation model for 33 languages that accepts instructions written in multiple languages.

Low-latency v4 speech model for expressive interactive voices.

A 24B regional-language model trained for Middle Eastern and South Asian language and cultural contexts.

A 7B translation model for 33 languages that accepts instructions written in multiple languages.

Hosted image model for text rendering and small visual details, with sampling-step and guidance controls.

A batch speech-to-text model with speaker labels, word timestamps and vocabulary hints for recordings.

GPT-5.4 Pro is the higher-compute GPT-5.4 release for difficult problems.

A deprecated GPT-5 reasoning model that uses additional computation through the Responses API.

Video generation and editing model with native audio and clips of 3–15 seconds at resolutions up to 4K.

An 8.38B embedding model for text queries and documents containing text, images or both.

Speech model for expressive narration, consistent character voices and multi-speaker dialogue.

Text embedding model for retrieval, clustering and personalization.

A deprecated o3 model with additional computation for difficult reasoning problems.

Multilingual speech model designed for short response delays.

A higher-compute GPT-5.2 model accessed through the Responses API.

Expressive speech model with audio tags and multi-speaker performance.

Video model that produces 3–10-second clips from text, images or video at up to 4K and 24 fps.

A preview speech-to-text model for recorded audio with timestamps and speaker diarization.

A public-preview speech model for latency-sensitive audio generation.

A public-preview speech model for expressive long-form audio.

A preview speech-generation model with voice cloning, emotion tags and controls for speaking pace.

Multilingual speech model for stable long-form narration.

Multimodal embedding model that represents text and images in a shared semantic space.

Speech, sound-effect and ambient-audio generator combining language-based control with acoustic diffusion.

A speech-generation model with native breath and laughter tags, 40 languages and HD or Turbo serving options.

Realtime v3 speech model for directed delivery in voice agents.

A speech-generation model with reference voices and nine-language output under a noncommercial weight license.

Deprecated multilingual Turbo speech release that preceded Flash.

GPT-Realtime-2.1 combines conversational audio with reasoning and tool use.

Deprecated English Turbo speech model with telephony-oriented output support.

English-only Flash speech model for interactive applications.

Clinical-audio fine-tune of Scribe v2 for terminology and dictated notes.

Batch speech-recognition model for long recordings, structured transcripts and multilingual audio.

A text-to-video and image-to-video model with a separate Fast option for image input.

An audio-capable Ministral 3B adaptation for transcription, translation and questions about recordings.

Runway’s video-editing model, using an edited keyframe to carry changes across clips and multiple shots.

Video generation model with synchronized audio, timed image keyframes and existing-video input.

Video super-resolution model with precise and creative modes that retains the original audio.

A video-generation model that accepts text prompts or starting images.

GPT-Transcribe turns recorded speech into text.

A video-generation model that accepts text, images and existing video.

Imagine Video 1.5 generates video from text or image input through a named API endpoint.

A smaller Imagine Video 1.5 model that generates video from text or image input.

Video generation model with sound and text, image and reference-to-video interfaces.

A public-preview transcription model with support for 60 languages.

A 25-language transcription model for live and recorded audio with speaker labels and vocabulary hints.

A 600M speech recognizer for streaming transcription across 40 language-locales.

An offline audio-to-text model for 40 languages that accepts recordings, instructions and optional context.

Compact Qwen3-Omni-based speech recognizer with streaming and offline inference.

Speech recognition model for streaming and recorded audio in 30 languages and 22 Chinese dialects.

Multilingual reranker for document text and serialized structured records.

Throughput-focused multilingual document reranker in the Rerank 4 family.

Multilingual search reranker for documents and semi-structured records, with a focus on ranking quality.

Video model for generation, editing and extension with text, image, video and audio references.

Faster Seedance 2.0 video variant with reference inputs and editing controls.

Cost-focused video model for generation, editing and extension in the Seedance 2.0 family.

Joint audio and video model that generates up to 30 seconds per request and can continue an existing video.

Video generation model with text, image, video and audio references and a faster Prime variant.

A deprecated hosted Whisper model for multilingual speech recognition, translation and language identification.

A Qwen derivative trained to change responses on political and social topics.

Pretrained sparse language model with 280 billion total parameters and 14 billion active per token.

A Sabiá model for Portuguese research, coding and long documents with adjustable reasoning and multi-step tool use.

Sound-generation model for sequences, environmental audio, vocals and musical effects.

An image-editing preview model launched on 27 October 2026.

Document parsing model with sparse-reference sliding-window attention for long documents.

Hy-SimulMT-30B-A3B supports incremental translation between Chinese and English.

An audio-to-text model with a Granite 4.0 1B Base backbone for speech recognition in six languages.

A streaming transcription model with turn detection, speaker diarization and configurable latency over WebSocket.

Multilingual medical recognition model for consultations that change language.

A classifier for Arabic language varieties and other languages that use Arabic script.

An Arabic streaming-transcription adaptation for dialects, Modern Standard Arabic and code-switched speech.

A 7B Falcon H1 Arabic adaptation for Emirati vocabulary, cultural references and conversation.

October 2026 speech-to-speech model for real-time conversations on Amazon Bedrock.

German-and-English reasoning model with 78.1 billion total parameters and 3.46 billion active per token.

LLM-jp-4.1 32B-A3B Thinking is a Japanese-English reasoning release with new SFT and preference training.

LLM-jp-4.1 33B Thinking is a Japanese-English reasoning release with new SFT and preference training.

LLM-jp-4.1 8B Thinking is a Japanese-English reasoning release with new SFT and preference training.

Vision 2.1 adds document capabilities for complex tables, form fields and Indian-language handwriting.

A document parser that converts page images into structured text.

Music-generation release focused on layered arrangements, instruments and vocal-led tracks.

Claude Mythos 5.1 gives verified organizations access to the Fable 5.1 model for defensive cybersecurity.

Speech recognition model for short voice-agent turns and spoken identifiers.

A 30B dense text model derived from Granite 4.1 Base with full, low-effort and direct-answer thinking modes.

A 3B dense text model derived from Granite 4.1 Base with full, low-effort and direct-answer thinking modes.

A closed-beta model for continuous video that responds to new instructions.

A policy classifier that checks text and images against plain-language rules and returns yes-or-no scores.

A listwise search ranker tuned for specialist documents and structured data.

Midjourney's current default image model.

A document-extraction model that returns text with paragraph boxes, structural labels and confidence scores.

An 8B navigation model that uses one RGB camera and text instructions to direct a robot through a scene.

Chunk embeddings that use surrounding document context and automatic splitting.

An OCR release that adds paragraph locations, structural block types and confidence scores to extracted text.

DeepSeek-OCR-based model for joint parsing of multiple pages in a single document-level pass.

Multilingual speech recognition model that follows language changes within a recording.

Claude Mythos 5 is the verified-access offering of Claude Fable 5 for defensive cybersecurity.

Audio-to-audio dubbing model that transfers a recorded performance into another language.

A 0.9-billion-parameter document vision-language model with improved training for difficult document regions.

Music model with section-level composition and local regeneration.

A compact multimodal encoder aligned with Jina's v5 Text Nano vectors.

A shared search encoder for text, images, audio, video and document files.

Luma’s image generation and editing model, with a higher-quality Max tier and reference-image controls.

Hosted general-purpose model derived from ERNIE 5.0 through elastic model selection and task-specific post-training.

An instruction-tuned 30B text model for retrieval-based answers, extraction, coding and tool use.

A pretrained 30B dense decoder for text generation, further training and fill-in-the-middle code completion.

An instruction-tuned 3B text model for retrieval-based answers, extraction, coding and tool use.

A pretrained 3B dense decoder for text generation, further training and fill-in-the-middle code completion.

An instruction-tuned 8B text model for retrieval-based answers, extraction, coding and tool use.

A pretrained 8B dense decoder for text generation, further training and fill-in-the-middle code completion.

A mixture-of-experts image model for text-to-image generation and instruction-based editing.

A fast image model with native high-definition generation.

A SAM update that uses Object Multiplex and shared memory to track several objects jointly.

GUI model that converts screenshots and instructions into operations, with thinking and non-thinking modes.

MiMo V2 Omni combined text, visual and speech understanding with tool calls and computer interaction.

MiMo V2 Pro was Xiaomi's text model for coding and multi-step agent work before V2.5 Pro.

An earlier speech-generation model combining an audio tokenizer with speech-text modeling and reinforcement learning.

The retired first public Alpha release of the V8 image model.

A sparse code model for Lean 4 proof engineering, with tool calling and image input.

A multilingual moderation model for long conversations, harmful-content detection and jailbreak checks.

A tiny six-layer model for studying the effect of training data.

A six-layer Pythia model that doubles the width of the 14M network.

Bilingual recognition model for Arabic dialects mixed with English speech.

Bilingual clinical recognition model for Arabic speech mixed with English medical terms.

A EuroBERT-based text encoder for search with a smaller memory footprint.

A multilingual text encoder with four task adapters and compact vector options.

General-purpose model for extended reasoning and complex visual workflows.

The Instruct release adds reasoning, prompt rewriting and image-to-image generation to HunyuanImage-3.0.

Downloadable multilingual embedding model with bidirectional attention.

Swedish clinical recognition model for specialist terms and multi-speaker consultations.

Compact dense language model for private document processing and instruction-based tasks.

Hybrid language model with 12 billion active parameters for document questions, extraction and structured text.

A Qwen3 derivative trained for competitive programming.

A hybrid Arabic language-model family with 3B, 7B and 34B versions.

Document recognition model that reorders visual tokens to follow page structure before it extracts text and layout.

An anime image model developed by Midjourney and Spellbrush.

An OCR model with HTML table output and support for handwritten forms, available through commercial document-processing services.

Danish clinical recognition model trained within the Nordic medical expansion.

Four-language recognition model for English, Mandarin, Malay and Tamil conversation.

A mathematical reasoning model developed with Hillclimb AI.

Norwegian clinical recognition model for live and recorded medical conversations.

A small vision-language model for questions about images and documents.

A Seed-OSS derivative post-trained across the Psyche network.

A separately trained control model for the Hermes 4.3 research.

An experimental Small adaptation for creative writing, stories, roleplay and character dialogue.

German clinical recognition model developed for compound terms and medical dialogue.

An image-editing model that adds reasoning and reflection to the Step1X-Edit pipeline.

Streaming recognition model with partial and committed transcripts.

First-generation Composer coding-agent model with a 200,000-token context.

Image generation and instruction-based editing model with native four-megapixel output.

A published model for multi-subject reference video.

Dutch clinical recognition model for medical conversations and documentation.

Finnish clinical recognition model in Speechmatics’ Nordic medical range.

French clinical recognition model for medical dialogue and specialist terms.

Spanish clinical recognition model for specialist vocabulary and patient conversations.

Small hybrid reasoning model trained for code, mathematics and structured answers.

A compact ranker that compares a query with several documents in one pass.

Pretrained hybrid-attention model that combines recurrent memory with selected global-attention layers.

Instruction-tuned hybrid-attention model designed to reduce memory use in long conversations.

A multimodal Magistral Medium update that expands the published context window to 128K.

English clinical recognition model trained for medical terminology and fast dialogue.

A 149M ModernBERT cross-encoder that scores English query-document pairs to refine search results.

A compact encoder that finds code from descriptions, examples and completion context.

A larger code-search encoder with adjustable vectors for technical retrieval.

A Qwen3-14B derivative with optional reasoning and structured output.

A Llama 3.1 70B derivative with optional reasoning and structured output.

A 149M English retrieval model that produces 768-dimensional vectors.

A 47M ModernBERT English retrieval encoder that produces 384-dimensional vectors.

Bilingual recognition model for Tagalog and English within the same recording.

Deprecated first Eleven Music release for songs and instrumentals.

A controlled suite of 6.9B models studying training-data filtering and safety.

The July 2025 hosted Magistral revision, with a 40K context window and reasoning-oriented tool use.

A 24B reasoning update with dedicated reasoning tokens, shorter simple answers and fewer generation loops.

End-to-end speech interpretation model for Chinese and English.

Document-aware chunk embeddings trained for chunk and document relevance.

Sound-effects model that uses text to describe sounds and optional voice input to control timing and energy.

The transcription-optimized Voxtral Mini variant used by the original hosted speech-to-text route.

Original instruction-tuned K2 checkpoint for tool use and coding without extended thinking mode.

A hosted software-engineering model for tool-driven repository exploration and changes across multiple files.

A 24B coding-agent update with broader prompt support and Mistral function-calling output.

An instruction-based image-editing model that also supports text-to-image generation.

Large hybrid mixture-of-experts model focused on answers grounded in supplied documents.

Earlier Mini release with improved source-grounded answers and instruction following.

A Japanese business language model with a 128K input capacity and support for task decomposition and tool selection.

Trillion-parameter pretrained model that introduced MuonClip training to the Kimi K2 family.

A visual-document encoder with both single-vector and token-vector retrieval.

Prompted voice-design model that creates and remixes synthetic voice profiles.

Revised visual-reasoning model with higher-resolution image input and improved video understanding.

An image-to-video model for animating a starting frame.

An earlier hosted video model with text and image input and 512P, 768P and 1080P output options.

Qwen2.5-72B adaptation trained to locate software faults and produce repository-level fixes.

The first hosted Magistral reasoning release, designed to work through multilingual and technical questions.

A downloadable 24B reasoning model trained from Small 3.1 with supervised examples and reinforcement learning.

A seven-billion-parameter Common Pile model trained for one trillion tokens.

A seven-billion-parameter Common Pile model trained for two trillion tokens.

An image-generation model that replaces U-Net with a diffusion transformer and uses selected artistic training data.

An updated MiMo 7B reasoning model trained with an expanded supervised dataset and a longer reinforcement-learning window.

The May 2025 document-extraction release with interleaved text, images and structured annotations.

A 24B text-only coding agent derived from Mistral Small 3.1 with its vision encoder removed.

A Falcon3-7B adaptation for formal Arabic and regional dialects.

Updated general embeddings distilled from larger retrieval models.

Smaller 3.5-series embeddings with flexible vector sizes.

Bilingual recognition model for Malay conversation mixed with English terms.

Bilingual recognition model trained for Mandarin and English code-switching in Southeast Asia.

Bilingual recognition model for Tamil and English speech in the same conversation.

A two-stage 3D generation model with separate geometry and texture models.

Dense reasoning model trained from Qwen3-8B-Base using reasoning examples generated by the larger R1-0528 model.

Pretrained audio-language model for adaptation to speech, sound and audio conversation tasks.

Instruction-tuned audio model that can answer in text or produce speech with its text response.

Step1X-Edit combines a multimodal language model that interprets edits with a diffusion transformer that generates the image.

Higher-detail image-generation model for complex scenes, released alongside Firefly Image 4.

AI21 and Enkrypt AI's alignment study based on Jamba Mini 1.5.

A 2B instruction model with optional structured reasoning and support for 12 languages.

An 8B instruction model with optional structured reasoning and support for 12 languages.

A 32-billion-parameter GLM release from April 2025.

A compact reasoning release from the April 2025 GLM family.

A relevance scorer for text and document images in multiple languages.

A hosted speech model with 24 languages, seven emotions and HD or Turbo serving options.

An experimental DeepHermes reinforcement-learning artifact.

An experimental DeepHermes reinforcement-learning artifact.

An experimental DeepHermes reinforcement-learning artifact.

A research model for predicting the direction of company financial metrics.

An experimental model trained to improve tool calls during reasoning.

Vision-language model for images, documents and video, built from Moonlight and the MoonViT encoder.

Original long-reasoning vision-language checkpoint, retained alongside the later June 2025 revision.

A classifier for refusals in assistant responses.

A video-generation model that animates a starting image from a text instruction.

A preview model with direct-answer and extended-reasoning modes.

A preview model with direct-answer and extended-reasoning modes.

398-billion-parameter Jamba release for long-document extraction and enterprise text tasks.

March 2025 Mini model for classification, document questions and structured generation.

Mistral’s first OCR model converts images and PDFs into ordered text and image content.

An 8B text model that adds controllable thinking to Granite 3.1 Instruct.

Deprecated first Scribe recognition model with speaker labels and audio-event tags.

Step-Video T2V is a 30B text-to-video model that can generate up to 204 frames.

Step-Video T2V Turbo is the inference-step-distilled version of Step-Video T2V.

An image-generation model with text prompts and a single portrait reference for character consistency.

A preview model with direct-answer and extended-reasoning modes.

Pretrained language model used to study Muon optimization at 16 billion total parameters.

Instruction-tuned Moonlight release for conversation and task prompts, with an 8,192-token context.

Compact Janus Pro model that pairs a SigLIP image encoder with a separate tokenizer for image generation.

The larger Janus Pro model combines image understanding and image generation through separate visual encoders and a shared language model.

Reasoning language model that combines supervised examples and reinforcement learning on a large mixture-of-experts foundation.

Research reasoning model trained with reinforcement learning directly from a pretrained base, without an initial supervised fine-tuning stage.

A long-context language model that combines lightning attention, softmax attention and a mixture-of-experts network.

A January 2025 video model designed for coherent motion, camera movement and ordered events.

A small language model that turns HTML into Markdown or schema-based JSON.

A later Codestral release that improves code completion speed through a revised architecture and tokenizer.

General embeddings with selectable dimensions and quantized output.

Multimodal reasoning research model used to study reinforcement learning and long-to-short reasoning transfer.

An open-weight Hermes assistant derived from Meta's Llama models.

Code embeddings with variable dimensions and quantization support.

HunyuanVideo is the original text-to-video release published in December 2024.

A 10B multilingual transformer language model created by extending the depth of Falcon3-7B-Base.

Falcon3 1B Instruct is a compact transformer text model for English, French, Spanish and Portuguese.

Falcon3 3B Instruct provides multilingual text generation in a 32K-context transformer.

A 7B transformer language model for English, French, Spanish and Portuguese with a 32K context.

A 7B instruction-tuned language model that continues Falcon Mamba training with a pure Mamba1 architecture.

A 10-billion-parameter base language model trained across distributed computers.

A multilingual text-and-image encoder with adjustable vector dimensions.

A 123B Large update focused on long-context use, system prompts and more accurate tool calls.

A 124B visual-language model that extends Mistral Large 2 with a one-billion-parameter image encoder.

Multimodal research model that combines autoregressive text generation with rectified flow to create images at 384 by 384 pixels.

Joint text and image embeddings for visual document retrieval.

A multilingual text classifier that evaluates raw text or the final message in a conversation.

First-generation Firefly video model that generates footage from text or image inputs.

Historical recognition generation that expanded self-supervised training and language coverage.

Original Janus model with separate image encoding paths; its early release required a tokenizer correction for classifier-free guidance.

The original compact Ministral release for local intelligence, tool routing and low-latency text tasks.

An 8B text model with alternating attention ranges and research-licensed instruction weights.

Second-generation reranker with a 16,000-token pair limit.

Smaller second-generation reranker with an 8,000-token pair limit.

A Japanese enterprise language model co-developed by Fujitsu and Cohere.

A multilingual text model with separate query and passage adapters.

Third-generation general text embeddings with a 32,000-token input limit.

Smaller third-generation encoder with 512-dimensional output.

A 22B Small release with function calling and downloadable instruction weights under research terms.

A NeMo-based visual-language model with a new 400M vision encoder and support for varied image sizes.

A small HTML-to-Markdown model for local page-content conversion.

The larger first-generation ReaderLM model for HTML-to-Markdown conversion.

319-million-parameter trained development checkpoint for work on the Jamba architecture.

A multilingual retriever that keeps separate vectors for document tokens.

First Large Jamba checkpoint, scaling the hybrid design to 398 billion total parameters.

Original instruction-tuned Mini model with tool calls, JSON output and long text input.

An open-weight Hermes assistant derived from Meta's Llama models.

An earlier image model with improvements to detail and quoted text.

A downloadable 3B code-specialist model fine-tuned from Granite's long-context base in July 2024.

A downloadable 8B code-specialist model fine-tuned from Granite's long-context base in July 2024.

A Mamba 2 coding model that uses a state-space architecture and provides Apache-licensed instruction weights.

A Mistral 7B adaptation for mathematical reasoning and scientific questions, with Apache-licensed weights.

A Hermes release trained for function calls and JSON responses.

A multilingual cross-encoder that scores query-document relevance.

A merged model with additional preference training.

Multilingual text embeddings for cross-language retrieval.

An English embedding model trained for both text search and image matching.

Financial text embeddings for filings, reports and related evidence.

A 22B coding model for instruction-based programming and completion between a code prefix and suffix.

Earlier general and multilingual relevance reranker.

A 7B release with a larger vocabulary and an instruction checkpoint that supports function calling.

A 20B instruction-tuned code model with an 8K context, released in May 2024.

A 34B instruction-tuned code model with an 8K context, released in May 2024.

Instruction-tuned embeddings for retrieval, clustering and classification.

Historical hosted instruction version of the original Jamba architecture.

A merged model with additional preference training.

An earlier compact cotomi model aimed at high request volume.

An earlier NEC model designed for more demanding Japanese text tasks.

Third-generation Firefly image model with image-detail, prompt-understanding and creative controls.

A four-layer English ranker distilled for fast relevance scoring.

A six-layer English ranker between the tiny and base sizes.

A compact encoder-decoder model trained on two trillion Pile tokens.

A 783-million-parameter encoder-decoder model trained from scratch for text-to-text research.

A 2.85-billion-parameter text-to-text model with an open training history.

An eleven-billion-parameter encoder-decoder model trained from scratch.

Legal text embeddings for long documents and case retrieval.

Original pretrained hybrid Mamba-Transformer checkpoint for model training and adaptation.

First lightweight Voyage reranker for retrieved document sets.

A Hermes release trained for function calls and JSON responses.

A model that generates instruction data from source text.

A research model that explores low-bit training.

The original English query-document relevance model in Jina's reranking family.

The original Mistral Large text model, with multilingual instruction following and native function calling.

A concise-answer prototype exposed in the original Le Chat beta rather than as a public weight release.

The February 2024 Small model introduced as a lower-latency counterpart to the original Mistral Large.

An earlier instruction-tuned Hermes model with downloadable weights.

An English token-level retriever built for longer documents.

A text-search encoder trained for Spanish and English.

A compact code-search encoder trained on source and documentation pairs.

A seven-billion-parameter RWKV chat model with recurrent inference.

Earlier API model family for long text and structured responses, with later vision-preview variants.

An anime model that responds to detailed descriptions of unusual subjects.

Earlier code retrieval encoder for natural-language and code queries.

A bilingual vector model for searching German and English text.

Mixtral 8x7B instruction model with a further direct preference optimization stage.

Supervised instruction tuning of Mixtral 8x7B, with ChatML conversation formatting.

A bilingual encoder for matching Chinese and English documents.

A Hermes release trained for function calls and JSON responses.

An earlier instruction-tuned Hermes model with downloadable weights.

A SOLAR base model extended through long-context training.

A SOLAR base model extended through long-context training.

An earlier instruction-tuned Hermes model with downloadable weights.

An earlier image model designed for more detailed prompts.

Bilingual recognition model that transcribes Spanish and English in one stream.

A 7B revision that increases context to 32K and removes the first release’s sliding-window attention.

The proprietary prototype behind Mistral’s original highest-capability hosted text endpoint.

Mistral’s first sparse expert text model, released as pretrained and instruction-tuned Apache weights.

A continued-pretraining model for longer Mistral sequences.

A conversation model trained with Nous's synthetic-data method.

A compact vision-language research release.

First public Voyage embedding model, now deprecated.

An English long-document encoder with permissively licensed weights.

A 34B mathematical language model adapted from Meta Code Llama.

A 7B mathematical language model adapted from Meta Code Llama.

Design model that combines layout, imagery and typography to create editable templates in Adobe Express.

Second-generation Firefly image model with reference-based style matching, photography controls and multilingual prompts.

A conversation model trained with Nous's synthetic-data method.

A Mistral instruction model published by Teknium.

A 33-million-parameter English encoder with long-document input.

Mistral’s first 7B Transformer release combines grouped-query attention with sliding-window attention.

A small code-completion model trained on synthetic programming exercises.

A conversation model trained with Nous's synthetic-data method.

This release adds PPO alignment to the bilingual 4B SFT model using Japanese-translated Anthropic HH-RLHF data.

An original-dataset Hermes fine-tune of Llama 2.

A tiny English sentence encoder for low-resource semantic search.

An original-dataset Hermes fine-tune of Llama 2.

Specialized question-answering model trained to use supplied evidence and identify missing answers.

A Jina-trained Falcon 40B adaptation using Code Alpaca examples.

A Jina-trained Falcon 7B adaptation for coding instructions.

A larger English embedding checkpoint with 1,024-dimensional output.

A small English vector model from Jina's first embedding release.

An original-dataset Hermes fine-tune of Llama 2.

A 14M Pythia checkpoint trained on deduplicated Pile text.

A 31M Pythia checkpoint trained on deduplicated Pile text.

A Llama 2 fine-tune focused on extended conversations.

Legacy classifier for estimating whether older ElevenLabs models generated an audio clip.

The base-size English model in Jina's original embedding family.

Small instruction-capable Jurassic-2 model for classification, extraction and short text.

The first-generation Hermes instruction model on Llama 13B.

Retired first multilingual speech model, covering English and seven additional languages.

A 24-layer Pythia network with a two-thousand-wide hidden state.

A 1.4B Pythia checkpoint trained on deduplicated Pile text.

The largest Pythia size, with 36 layers and open training checkpoints.

A 12B Pythia checkpoint trained on deduplicated Pile text.

A twelve-layer open model with checkpoints across its training run.

A 160M Pythia checkpoint trained on deduplicated Pile text.

A wide, sixteen-layer Pythia model trained with bfloat16 precision.

A 1B Pythia checkpoint trained on deduplicated Pile text.

A 32-layer Pythia model with about 2.52 billion non-embedding parameters.

A 2.8B Pythia checkpoint trained on deduplicated Pile text.

A 24-layer research model with a 1,024-wide hidden state.

A 410M Pythia checkpoint trained on deduplicated Pile text.

A 32-layer open model for research at a six-billion-parameter scale.

A 6.9B Pythia checkpoint trained on deduplicated Pile text.

The smallest size in the original eight-size Pythia research suite.

A 70M Pythia checkpoint trained on deduplicated Pile text.

An anime illustration model with deliberately strong visual styling.

Original Firefly model for image generation, text effects and Photoshop Generative Fill.

Historical middle Jurassic-2 base model from the March 2023 release.

Historical largest Jurassic-2 base model from the March 2023 release.

Historical smallest Jurassic-2 base model from the March 2023 release.

Middle Jurassic-2 tier for text generation, summaries and information extraction.

Largest instruction-capable Jurassic-2 tier for complex text comprehension and generation.

Historical recognition generation with self-supervised audio representations and separate decoding components.

A legacy model with a photographic default appearance.

Historical voice-generation model controlled by age, gender and accent selections.

Language-specific FastPitch and HiFi-GAN speech synthesis models, with public checkpoint archives.

An early instruction fine-tune of Vicuna 13B.

A conversation model trained with Nous's synthetic-data method.

An early vision-language model with function-call training.

A Llama 2 fine-tune focused on extended conversations.

A Teknium instruction model trained on a public Hermes dataset.

A coding-enriched continuation of OpenHermes 2.

A Teknium instruction model trained on a public Hermes dataset.

An experimental instruction fine-tune of WizardCoder.

A Llama 2 model with continued training for longer context.

A Llama 2 model with continued training for longer context.

A long-context research extension of Llama 2 70B.

A Llama 2 model with continued training for longer context.

A Llama 2 model with continued training for longer context.

A continued-pretraining model for longer Mistral sequences.

The first Niji anime image model.

An image model built on Midjourney's redesigned architecture.

A Stable Diffusion adaptation for Japanese prompts and image concepts.

Japanese CLIP ViT-B/16 learns shared image and Japanese-text representations through contrastive image-text training.

Japanese CLOOB ViT-B/16 learns shared image and Japanese-text representations through Contrastive Leave One Out Boost training.

Middle-sized Jurassic-1 text model introduced between Large and Jumbo.

A 20-billion-parameter Pile model from EleutherAI distributed training research.

Historical self-supervised recognition release trained on a much larger audio collection.

A Japanese masked-language foundation encoder with specific tokenizer and position-ID requirements.

178-billion-parameter Jurassic-1 model with a wider, shallower Transformer design.

Historical 7.5-billion-parameter English text model with a large phrase-aware vocabulary.

A six-billion-parameter text model built with Mesh Transformer JAX.

An early open GPT-style model trained on 380 billion tokens.

The largest original GPT-Neo checkpoint, with 32 Transformer layers.

FinBERT Base Cased is a Finnish masked-language encoder with a dedicated WordPiece vocabulary.

FinBERT Base Uncased is a Finnish masked-language encoder with a dedicated WordPiece vocabulary.

A performance-transfer model that applies speech, expression and movement to a character image or video.

4.5-billion-parameter instruction-tuned foundation model released under Apache 2.0.

Ǎguila 7B is BSC's earlier Catalan, Spanish and English language adaptation of TII's Falcon 7B.

Hindi instruction-tuned OpenHathi adaptation trained on AI4Bharat's IndicInstruct collection.

Large multilingual base model that underpins later 40B instruction-following and tool-calling adaptations.

Non-thinking 40-billion-parameter model trained for function calling.

ALIA 40B FC 2606 adds explicit reasoning to BSC's function-calling model line.

Reasoning and function-calling model with 40 billion parameters and access-controlled weights.

ALIA 40B Instruct 2601 is a long-context conversational release from BSC's 40B language model family.

ALIA 40B Instruct 2605 is a distinct instruction-tuned release aimed at more precise prompt and output-format control.

ALIA 40B Instruct 2606 combines long-context instruction following with BSC's helpfulness and safety post-training.

The Arabic and English model introduced with HUMAIN Chat, building on the earlier ALLaM research program at NCAI and SDAIA.

NCAI's downloadable Arabic and English 7B model, now published through HUMAIN, with a 4,096-token context and documented revision caveats.

Second-generation Nova speech-to-speech model for conversations on Amazon Bedrock.

Retired Amazon image-generation model whose endpoint reached end of life on 30 September 2026.

Embedding model that maps multiple content types into a shared vector space.

Retired text-and-image-to-video model whose endpoint closed on 30 September 2026.

An 8B Apertus instruction model adapted for Southeast Asian text tasks with published post-training data.

Arabic Stable LM 2 Base 1.6B adapts Stable LM 2 1.6B with more than 100 billion tokens of Arabic text.

Arabic Stable LM 2 Chat 1.6B is the instruction-tuned companion to the Arabic base.

A text encoder with fourfold pooling trained for use with a Llama decoder.

A text encoder with eightfold pooling trained for the Llama 3.1 8B decoder.

A Llama-derived text encoder with eightfold pooling trained for the Mistral 7B decoder.

A text encoder with eightfold pooling trained with both Llama 3.1 8B and Mistral 7B decoders.

70-billion-parameter merged model based on Meta's Llama 3.1 70B Instruct.

AuK is a 1.5B speech generation and editing model trained on millions of hours of audio.

A phone-control checkpoint based on GLM-4.1V-9B.

The multilingual published AutoGLM phone checkpoint.

Multilingual text model with 32 billion parameters and support for 23 languages.

Eight-billion-parameter multilingual text checkpoint in the Aya Expanse family.

A 32-billion-parameter vision-language model with image and text input and text output in 23 languages.

Eight-billion-parameter research model for image and text understanding across 23 languages.

BAGEL unifies text and image understanding, image generation and image editing.

BFS-Prover V2 32B generates Lean 4 proof tactics from a formal tactic state.

BFS-Prover V2 7B is the smaller Lean 4 tactic generator in the V2 release.

This register classifier adapts BGE-M3 RetroMAE and publishes a specific trained cross-validation fold.

This bundle contains 17 trained classifiers for subregister labels over precomputed BGE-style vectors.

Research speech recognizer adapting a 1.2-billion-parameter Canary architecture to Dehvali Bhili.

600-million-parameter IndicConformer adaptation trained on about 200 hours of Bhili speech.

Streaming Bhili speech recognizer derived from Nemotron Conformer with selectable decoding latency.

Sarvam-derived LoRA model for Dehvali-Bhili-to-Marathi translation with an agricultural focus.

Sarvam-derived LoRA model for Marathi-to-Dehvali-Bhili translation with agricultural training data.

Dehvali Bhili speech model fine-tuned from Indic Parler-TTS on ten hours of studio recordings.

An English-Japanese GPT-NeoX language model with 3.8B parameters.

An English-Japanese 3.8B language model with an 8,192-token sequence length.

This visual-language release joins rinna's 3.8B bilingual GPT-NeoX model with BLIP-2 components and a projection layer.

The supervised instruction release adapts Bilingual GPT-NeoX 4B using English and Japanese instruction data.

BLUUMI adds Finnish capability to BigScience's multilingual BLOOM 176B model through continued pretraining.

Qwen3.5-2B adaptation for reading English Braille from images, trained on synthetic examples.

A classifier for prompt-injection attempts in browser content.

An earlier Bulbul text-to-speech model with a distinct set of speech controls.

A text-to-speech model with more than 30 named voices across ten Indian languages and English.

A text-to-speech model with a persona-based catalog of 222 voice IDs.

One-billion-parameter punctuation-restoration model for English and 22 Indian languages.

270-million-parameter punctuation-restoration model for English and 22 Indian languages.

A causal video autoencoder with one-frame temporal compression and a 16-by-16-pixel spatial grid.

A causal video autoencoder with one-frame temporal compression and a 48-by-48-pixel spatial grid.

A causal video autoencoder with two-frame temporal compression and a 16-by-16-pixel spatial grid.

Four separately trained world models that predict video from context frames and actions in distinct environments.

A multi-resolution CADRE autoencoder with two-frame and 16-by-16-pixel compression.

A causal video autoencoder with two-frame temporal compression and a 32-by-32-pixel spatial grid.

A multi-resolution CADRE autoencoder with two-frame and 32-by-32-pixel compression.

A causal video autoencoder with four-frame temporal compression and a 16-by-16-pixel spatial grid.

A multi-resolution CADRE autoencoder with four-frame and 16-by-16-pixel compression.

A causal video autoencoder with four-frame temporal compression and a 24-by-24-pixel spatial grid.

A causal video autoencoder with eight-frame temporal compression and a 16-by-16-pixel spatial grid.

A multi-resolution CADRE autoencoder with eight-frame and 16-by-16-pixel compression.

Four inverse-dynamics models that estimate actions from generated frames to evaluate CADRE world models.

A vision-language model that links Helium 1 with a Qwen 2.5 VL encoder through cross-attention.

A Helium vision-language model that shares weights between self-attention and cross-attention.

A Qwen 2.5 VL 3B Instruct adaptation that replaces image-token insertion with cross-attention.

A CASA Qwen checkpoint fine-tuned on LiveWhisperX-526K for live video commentary.

A Qwen-based vision-language adaptation with shared self-attention and cross-attention weights.

Chat Latest is a moving alias for the Instant model used in ChatGPT.

chatgpt-image-latest is a deprecated alias for an image snapshot previously used in ChatGPT.

A restricted-access general-purpose model with cybersecurity capabilities, introduced for Project Glasswing.

A hosted video model with text, image and endpoint-frame inputs.

A five-billion-parameter video model with text-to-video and image-to-video checkpoints.

A downloadable image generator with explicit precision limits.

Two-billion-parameter speech recognition model for 14 languages, with Apache 2.0 weights.

Two-billion-parameter Arabic speech recognizer with regional dialect and Arabic-English code-switching support.

A 111-billion-parameter reasoning model with controlled thinking and a 256,000-token context across 23 languages.

Dedicated translation model with 111 billion parameters and support for 23 languages.

Vision-language model for documents, charts, tables and visual questions, with up to 20 images per request.

Cursor’s coding-agent model, built on Moonshot’s Kimi K2.5.

BSC's LoS Conformer adapts NVIDIA's Spanish Conformer Transducer to four languages of Spain.

This BSC LoS Conformer release adds trained punctuation to multilingual speech recognition.

ConMamba Small Catalan is BSC's Catalan speech recogniser built with convolution-augmented Mamba state-space layers.

ConMamba Small Spanish is BSC's Spanish speech-recognition release using a convolution and state-space architecture.

The faster option in NEC's second-generation cotomi release.

The higher-accuracy option in NEC's second-generation cotomi release.

An iterative neural model trained for ImageNet image classification.

An iterative neural model trained on 39 by 39 mazes to predict movement sequences.

Character-level forced-alignment model for placing supplied text on an audio timeline.

A custom embedding model used for semantic code search.

Cursor’s custom model for predicting edits and the next editing location.

A 1.92B decision model for text, images and video.

An earlier 3.1B decision model based on LFM2.5-VL-3B.

A 2B-class decision model that adds audio input to Liquid's decision line.

A 587M experimental decision model.

A set of fastText classifiers that filters Common Crawl text in the 24 official EU languages.

Daybreak Blue is a deprecated alias for general-purpose models used in defensive security work.

Daybreak Red is a deprecated cybersecurity alias.

Decision model that converts business text into structured classifications, scores and probabilities.

Hosted text and document translation model with terminology and context controls.

Real-time speech translation model for captions and spoken output.

Writing model for grammar corrections, rephrasing and tone adjustments.

A Japanese reasoning model combining Bakeneko with a DeepSeek-R1-Distill-Qwen-32B chat vector and ORPO.

An Apache-licensed 24B coding-agent model that adds image input to repository and tool workflows.

Nari's original English dialogue speech model, with speaker tags, audio prompting and downloadable 1.6B weights.

The June 2025 Dia revision with documented Transformers integration, English dialogue generation and audio-conditioned voices.

The smaller Dia2 checkpoint for English streaming dialogue, with reference voices and a two-minute generation limit.

The larger Dia2 streaming dialogue model, with a wider decoder, audio-prefix conditioning and word-timing output.

A Gemma-derived chat checkpoint trained with a discovered preference objective.

A hypernetwork that generates Gemma 2 2B instruction-model LoRA adapters from document content.

A hypernetwork that generates Mistral 7B Instruct v0.2 LoRA adapters from document content.

A hypernetwork that generates Qwen3 4B Instruct 2507 LoRA adapters from document content.

Multimodal embedding model for mixed text, image and video inputs, with dense and multi-vector output.

A Continuous DDPM model trained on natural-biome data to generate Minecraft voxel chunks.

A Discrete MD4 model trained on natural-biome data to generate Minecraft voxel chunks.

A Discrete MD4 model trained on natural-biome data to generate Minecraft voxel chunks.

Published dense and expert-model experiments for studying partial weight reinitialization during upcycling.

ECCO-BERT is an English encoder trained on digitized eighteenth-century documents from the United Kingdom.

Retired original English speech model that preceded the multilingual family.

Text, image and mixed-document embedding model with selectable vector dimensions and a 128K context.