Sakana AI
Sakana AI is a Tokyo research company founded by David Ha, Ren Ito and Llion Jones. Its work includes model merging, distillation and new neural architectures. Namazu is its Japanese business model; many earlier experiments have downloadable research weights.
Models & tools
- ctm-imagenet
Sakana AI · Continuous Thought Machines
An iterative neural model trained for ImageNet image classification.
Launch date not confirmedResearch releaseDownloadable - ctm-maze-large
Sakana AI · Continuous Thought Machines
An iterative neural model trained on 39 by 39 mazes to predict movement sequences.
Launch date not confirmedResearch releaseDownloadable - DiscoPOP-zephyr-7b-gemma
Sakana AI · DiscoPOP
A Gemma-derived chat checkpoint trained with a discovered preference objective.
Launch date not confirmedResearch releaseDownloadable - Doc-to-LoRA for Gemma 2 2B
Sakana AI · Doc-to-LoRA
A hypernetwork that generates Gemma 2 2B instruction-model LoRA adapters from document content.
Launch date not confirmedResearch releaseDownloadable - Doc-to-LoRA for Mistral 7B
Sakana AI · Doc-to-LoRA
A hypernetwork that generates Mistral 7B Instruct v0.2 LoRA adapters from document content.
Launch date not confirmedResearch releaseDownloadable - Doc-to-LoRA for Qwen3 4B
Sakana AI · Doc-to-LoRA
A hypernetwork that generates Qwen3 4B Instruct 2507 LoRA adapters from document content.
Launch date not confirmedResearch releaseDownloadable - DreamCubedCP2
Sakana AI · DreamCubed
A Continuous DDPM model trained on natural-biome data to generate Minecraft voxel chunks.
Launch date not confirmedResearch releaseDownloadable - DreamCubedDP2
Sakana AI · DreamCubed
A Discrete MD4 model trained on natural-biome data to generate Minecraft voxel chunks.
Launch date not confirmedResearch releaseDownloadable - DreamCubedDP4
Sakana AI · DreamCubed
A Discrete MD4 model trained on natural-biome data to generate Minecraft voxel chunks.
Launch date not confirmedResearch releaseDownloadable - Evo-Nishikie-v1
Sakana AI · Evo-Ukiyoe
A ControlNet model trained with Evo-Ukiyoe to add color to ukiyo-e images.
Launch date not confirmedResearch releaseDownloadable - Evo-Ukiyoe-v1
Sakana AI · Evo-Ukiyoe
An EvoSDXL-JP adaptation with a trained LoRA for Japanese woodblock-print imagery.
Launch date not confirmedResearch releaseDownloadable - EvoLLM-JP-A-v1-7B
Sakana AI · EvoLLM-JP
A Japanese evolutionary merge using Arithmo2 rather than WizardMath.
Launch date not confirmedResearch releaseDownloadable - EvoLLM-JP-v1-10B
Sakana AI · EvoLLM-JP
A larger Japanese merge derived from three 7B source models.
Launch date not confirmedResearch releaseDownloadable - EvoLLM-JP-v1-7B
Sakana AI · EvoLLM-JP
A Japanese language-model merge that combines language and mathematics skills.
Launch date not confirmedResearch releaseDownloadable - EvoSDXL-JP-v1
Sakana AI · EvoSDXL-JP
An SDXL image-generation model combining Japanese prompt understanding with SDXL-Lightning.
Launch date not confirmedResearch releaseDownloadable - EvoVLM-JP-v1-7B
Sakana AI · EvoVLM-JP
A vision-language model that combines Japanese language understanding with LLaVA through evolutionary merging.
Launch date not confirmedResearch releaseDownloadable - fwpkm-l12-fa-pkm2_6_10
Sakana AI · FwPKM
A 12-layer research model combining full attention with static product-key memory.
Launch date not confirmedResearch releaseDownloadable A 12-layer research model combining full attention with static and fast-weight product-key memory.
Launch date not confirmedResearch releaseDownloadableA 12-layer research model combining gated deltanet, full attention and static product-key memory.
Launch date not confirmedResearch releaseDownloadableA 12-layer research model combining gated deltanet, full attention and static and fast-weight product-key memory.
Launch date not confirmedResearch releaseDownloadableA 12-layer research model combining gated deltanet, sliding-window attention and static product-key memory.
Launch date not confirmedResearch releaseDownloadableA 12-layer research model combining gated deltanet, sliding-window attention and static and fast-weight product-key memory.
Launch date not confirmedResearch releaseDownloadable- fwpkm-l12-gdn-pkm2_6_10
Sakana AI · FwPKM
A 12-layer research model combining gated deltanet with static product-key memory.
Launch date not confirmedResearch releaseDownloadable A 12-layer research model combining gated deltanet with static and fast-weight product-key memory.
Launch date not confirmedResearch releaseDownloadable- gpt-oss-120b-sft-aimo3-fishmath
Sakana AI · FishMath
A math-specialized fine-tune of OpenAI's gpt-oss-120b.
Launch date not confirmedResearch releaseDownloadable A trained Moshi extension that uses text guidance during spoken dialogue.
Launch date not confirmedResearch releaseDownloadable- Llama-2-7b-hf-DroPE
Sakana AI · DroPE
A recalibrated Llama 2 model with positional embeddings removed.
Launch date not confirmedResearch releaseDownloadable Coding expert used by the CycleQD project.
Launch date not confirmedResearch releaseDownloadable- Llama-3-8B-Instruct-CycleQD-CS
Sakana AI · CycleQD
A merged Llama 3 model for coding, database and operating-system tasks.
Launch date not confirmedResearch releaseDownloadable - Llama-3-8B-Instruct-DB-Expert
Sakana AI · CycleQD
Database expert used by the CycleQD project.
Launch date not confirmedResearch releaseDownloadable - Llama-3-8B-Instruct-OS-Expert
Sakana AI · CycleQD
Operating-system expert used by the CycleQD project.
Launch date not confirmedResearch releaseDownloadable - Llama-3-EvoVLM-JP-v2
Sakana AI · EvoVLM-JP
A merged Llama 3 vision-language model for interleaved images and Japanese text.
Launch date not confirmedResearch releaseDownloadable - Llama-3-Karamaru-v1
Sakana AI · Karamaru
A Japanese Llama 3 adaptation trained on historical text to generate Edo-period-style language.
Launch date not confirmedResearch releaseDownloadable - Llama-3.1-Namazu-405B
Sakana AI · Namazu
An earlier Namazu alpha adaptation of Meta Llama 3.1 405B.
Launch date not confirmedAccess not confirmed A Vision Transformer classifier that identifies individual handwritten historical Japanese characters.
Launch date not confirmedResearch releaseDownloadableAn earlier Namazu alpha adaptation of DeepSeek-V3.1-Terminus.
Launch date not confirmedAccess not confirmed- Namazu-gpt-oss-120B
Sakana AI · Namazu
An earlier Namazu alpha adaptation of OpenAI gpt-oss-120b.
Launch date not confirmedAccess not confirmed An OLMo 2 derivative trained with context repositioning.
Launch date not confirmedResearch releaseDownloadableAn OLMo 2 derivative trained with context repositioning.
Launch date not confirmedResearch releaseDownloadableA 32B reasoning student trained from a reinforcement-learned teacher.
Launch date not confirmedResearch releaseDownloadableA 7B reasoning student trained from a reinforcement-learned teacher.
Launch date not confirmedResearch releaseDownloadable- Sakana Namazu v1.0
Sakana AI · Namazu
A Japanese business model adapted from Kimi K2.6 that accepts text, images and files.
Launch date not confirmedAPI - SparseLM0.5B
Sakana AI · SparseLM
A 0.5B sparse language-model research checkpoint.
Launch date not confirmedResearch releaseDownloadable - SparseLM1.5B
Sakana AI · SparseLM
A 1.5B sparse language-model research checkpoint.
Launch date not confirmedResearch releaseDownloadable - SparseLM1B
Sakana AI · SparseLM
A 1B sparse language-model research checkpoint.
Launch date not confirmedResearch releaseDownloadable - SparseLM2B
Sakana AI · SparseLM
A 2B sparse language-model research checkpoint.
Launch date not confirmedResearch releaseDownloadable - TAID-LLM-1.5B
Sakana AI · TAID
An English language model distilled from a larger Qwen2 teacher.
Launch date not confirmedResearch releaseDownloadable - TAID-VLM-2B
Sakana AI · TAID
A compact vision-language model distilled from InternVL2-8B.
Launch date not confirmedResearch releaseDownloadable - Text-to-LoRA for Gemma 2 2B
Sakana AI · Text-to-LoRA
A hypernetwork that generates Gemma 2 2B instruction-model LoRA adapters from task descriptions.
Launch date not confirmedResearch releaseDownloadable - Text-to-LoRA for Llama 3.1 8B
Sakana AI · Text-to-LoRA
A hypernetwork that generates Llama 3.1 8B Instruct LoRA adapters from task descriptions.
Launch date not confirmedResearch releaseDownloadable - Text-to-LoRA for Mistral 7B
Sakana AI · Text-to-LoRA
A hypernetwork that generates Mistral 7B Instruct v0.2 LoRA adapters from task descriptions.
Launch date not confirmedResearch releaseDownloadable - TinySwallow-1.5B
Sakana AI · TinySwallow
A compact Japanese model co-developed with the Swallow team.
Launch date not confirmedResearch releaseDownloadable - TinySwallow-1.5B-Instruct
Sakana AI · TinySwallow
A Japanese instruction-tuned language model derived from the Qwen-based TinySwallow model.
Launch date not confirmedResearch releaseDownloadable