LLM-jp-4 32B-A3B Thinking

LLM-jp-4 32B-A3B Thinking adds supervised fine-tuning and direct preference optimization to the corresponding LLM-jp-4 base model.

Research release

Reasoning trained without reinforcement learning

The card specifies SFT followed by DPO, without reinforcement learning. It links the instruction and size-specific preference datasets. Its Harmony-style chat format requires the model's own tokenizer; the openai-harmony tokenizer is not interchangeable.LLM-jp-4 32B-A3B Thinking official model card

Measure reasoning output as part of serving

Practical guidance. Active parameters describe computation, not the full weight storage requirement. Plan memory for the complete model and its serving state.

GGUF is a distribution of this model

The project publishes an official GGUF repository for this Thinking release. It is a file-format and quantization distribution of the model, not an additional pretrained model.LLM-jp-4 32B-A3B Thinking official GGUF release

Specifications & access

LLM-jp-4 32B-A3B Thinking specifications
DetailValue and conditions
StatusResearch release
AccessDownloadable
Model identifiersllm-jp-4-32b-a3b-thinkingllm-jp-4-32b-a3b-thinking-gguf
Context window65,536 tokens

Maximum sequence length stated in the model card.

LLM-jp-4 32B-A3B Thinking official model card
ArchitectureMixture of experts; 128 routed experts, eight active per token.LLM-jp-4 32B-A3B Thinking official model card
Parameters32,139,028,992 total; 3,827,476,992 active parameters.LLM-jp-4 32B-A3B Thinking official model card
LicenseApache 2.0; downloadable research weights.LLM-jp-4 32B-A3B Thinking official model card

Related models

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

    Launch date Research releaseDownloadable
  • LLM-jp-4 32B-A3B Base provides Japanese and English text pretraining weights with a 65,536-token context.

    Launch date not confirmedResearch releaseDownloadable