Model library / llm-jp
llm-jp-4.1-32b-a3b-thinking is a 32-billion-parameter MoE Japanese language model fine-tuned with SFT and DPO.
Mixture of experts, count not published
Context length in tokens, as published
Weights at the released precision
Licence declared on the repository
Overview
The model is part of the llm-jp-4.1 series created by the Research and Development Center for Large Language Models at the National Institute of Informatics. It is released as a post-trained checkpoint that has undergone supervised fine-tuning and direct preference optimisation, without reinforcement learning, and is intended for general Japanese language tasks.
In practice the model comprises 32.1 billion parameters in a Qwen3MoeForCausalLM architecture with mixture-of-experts (expert count not disclosed). It supports a 65,536-token context window, is distributed in bfloat16 format, and is licensed under Apache-2.0, allowing commercial and research use.
Specification
| Repository | llm-jp/llm-jp-4.1-32b-a3b-thinking |
| Publisher | llm-jp |
| Published | 16 September 2026 |
| Architecture | Qwen3MoeForCausalLM |
| Model type | qwen3_moe |
| Parameters | 32,139,028,992 |
| Layers | 32 |
| Hidden size | 2,560 |
| Attention heads | 40, 4 key/value heads |
| Context length | 65,536 tokens |
| Vocabulary | 196,608 tokens |
| Weight format | bfloat16 |
| Licence | apache-2.0 |
Memory
Parameter count multiplied by the bytes each format uses. The key-value cache comes on top.
| Precision | Weights | Notes |
|---|---|---|
| bf16 | 59.9 GiB | Full precision as released |
| fp8 | 29.9 GiB | 8-bit, near-lossless on most models |
| int4 | 15.0 GiB | 4-bit, smallest footprint |
The cache costs 64 KiB per token at 16-bit, so the full 65,536-token context of one request needs about 4.0 GiB. Concurrency multiplies that number, not the weights.
Hardware
At the released precision, with 90% of the card given to the server and 2.5 GiB kept for the runtime. The context column is what is left for the cache on that setup.
| GPU | Cards needed | Free for cache | Context it holds |
|---|---|---|---|
| RTX 4090 24GB | 4 | 16.5 GiB | the full 65,536 tokens |
| RTX 5090 32GB | 4 | 45.3 GiB | the full 65,536 tokens |
| L40S 48GB | 2 | 21.5 GiB | the full 65,536 tokens |
| A100 80GB | 1 | 9.6 GiB | the full 65,536 tokens |
| H100 80GB | 1 | 9.6 GiB | the full 65,536 tokens |
| RTX PRO 6000 96GB | 1 | 24.0 GiB | the full 65,536 tokens |
| H200 141GB | 1 | 64.5 GiB | the full 65,536 tokens |
Serving
A vLLM launch line for the shape above. Check the model card for a runtime the publisher recommends.
vllm serve llm-jp/llm-jp-4.1-32b-a3b-thinking \ --max-model-len 65536 \ --gpu-memory-utilization 0.90
Questions
The weights take about 60 GiB at the released precision. That fits on one A100 80GB, which leaves roughly 10 GiB for the key-value cache.
The repository declares apache-2.0. Read the licence text before commercial use: the name alone does not tell you what is allowed.
The configuration allows 65,536 tokens. Whether the whole window is usable depends on the memory left for the cache, which the hardware table works out per card.
This page is a reference, and Aiqre does not serve this model today. We deploy open-weight models on dedicated EU hardware on request, so if you want an OpenAI-compatible endpoint for it, email hello@aiqre.com with the model name and roughly what volume you expect.
Also in the library
llm-jp-4.1-33b-thinking is a 33.2 billion-parameter Japanese Llama model fine-tuned with SFT and DPO for reasoning tasks.
Hy-Memory-30B-A3B is a Tencent-developed long-term-memory writer model designed to replace LLMs in Mem0, MemOS and Hy-Memory fram…
Granite Switch 4.1 30B is a modular LLM combining a 30-billion base, decision and guardian adapters, and an English speech-to-tex…
Sabi-Noema-MLA is a 27.4 B-parameter agentic model derived from Qwen3.8-27B, using multi-head latent attention to reduce attentio…
Hosting
Tell us the model and roughly what volume you expect. We reply with a price and an OpenAI-compatible endpoint, with your prompts never stored.