LLM-jp-4.1 32B-A3B Thinking
LLM-jp-4.1 32B-A3B Thinking updates post-training for the 32B-A3B architecture. It retains a 65,536-token context and publishes the SFT and model-specific DPO datasets.
Research release
The revision changes post-training
The card specifies SFT and DPO without reinforcement learning. It records 11.7 trillion pretraining and mid-training tokens. Its chat template supports Harmony-style messages, but requires the included tokenizer. Some pretraining material remains unpublished because of licensing limits.LLM-jp-4.1 32B-A3B Thinking official model card
Keep active and total size separate
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 also publishes GGUF files. Its reviewed card says llama.cpp needs the project's tokenizer-handling fork; an unmodified upstream build can fail to parse chat messages. These files remain part of the same trained release.LLM-jp-4.1 32B-A3B Thinking official GGUF release
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Launch date | |
| Access | Downloadable |
| Model identifiers | llm-jp-4.1-32b-a3b-thinkingllm-jp-4.1-32b-a3b-thinking-gguf |
| Context window | 65,536 tokens Maximum sequence length stated in the model card. LLM-jp-4.1 32B-A3B Thinking official model card |
| Parameters | 32,139,028,992 total; 3,827,476,992 active parameters.LLM-jp-4.1 32B-A3B Thinking official model card |
| Architecture | Mixture of experts; 128 routed experts, eight active per token.LLM-jp-4.1 32B-A3B Thinking official model card |
| License | Apache 2.0; research weights, not a provider-operated API.LLM-jp-4.1 32B-A3B Thinking official model card |
Related models
- LLM-jp-4 32B-A3B Thinking
LLM-jp / NII LLMC ยท LLM-jp-4
LLM-jp-4 32B-A3B Thinking provides Japanese and English text reasoning weights with a 65,536-token context.
Launch date not confirmedResearch releaseDownloadable