Japan's National Institute of Informatics released LLM-jp-4-33b-thinking — a 33B dense, Apache 2.0 reasoning model, one of three variants in the LLM-jp-4 series alongside an 8B dense and a 32B-A3B MoE (128 routed experts, 8 active, 3.8B active parameters). All three share a 65,536-token context and a training run of 11.7 trillion tokens across pretraining and mid-training. What makes it worth a post isn't the size — 33B is modest by this month's standards — it's who built it and how.
A public research institute, not a company
NII isn't a lab racing for market share; it's Japan's national institute for informatics research, and LLM-jp is closer to public infrastructure than a product line. That shows up in the release itself: pretraining, mid-training, SFT, and DPO datasets are all publicly linked from the model card, with only the portions genuinely blocked by third-party licensing withheld. That's a rare level of pipeline transparency — the same standard we credited Soofi S for out of Germany, now matched from Japan's academic sector rather than a startup. It also lands as a direct counterpoint to Sakana AI's Namazu, Japan's other answer to sovereign AI, which explicitly skips pretraining and post-trains other labs' open weights instead. NII is the "actually train one from scratch" half of the same national conversation — proof the cheaper strategy isn't the only one Japan is running.
Reasoning without reinforcement learning
The model card states it plainly, almost in passing: post-training uses supervised fine-tuning and direct preference optimization, without reinforcement learning. That's the detail worth sitting with. Nearly every "thinking" or reasoning-capable release this month — Qwen3.8-Max, DeepSeek-V4-Pro, GLM-5.3 — has built its reasoning behavior through an RL stage, several of them shipping mandatory always-on reasoning with no way to disable it as a direct consequence. LLM-jp-4-33b-thinking gets a working reasoning mode from SFT and DPO alone. Whether that's a resource constraint, a methodological preference, or a genuine finding that RL isn't necessary for reasoning behavior at this capability tier, the card doesn't say. But it's a clean natural experiment sitting right next to a month of RL-heavy releases, and a useful data point for anyone trying to isolate what RL is actually buying the labs that use it.
What the benchmark table shows
NII evaluates with an LLM-as-judge setup using gpt-5.4-2026-03-05. It states outright that switching judges from the gpt-4o used for the prior llm-jp-3 generation makes scores stricter and not directly comparable across generations — the kind of methodology disclosure worth crediting on its own. Against that judge, NII reports results on MT-Bench (JA/EN), AnswerCarefully (a 336-question Japanese safety benchmark), and llm-jp-instructions (400 single-turn questions). LLM-jp-4-33b-thinking beats gpt-oss-20b — OpenAI's own open-weight release — on every metric at comparable reasoning effort settings: MT-Bench Japanese 8.00 vs. 7.33, English 8.24 vs. 7.85, AnswerCarefully 3.79 vs. 3.55, and llm-jp-instructions 3.79 vs. 3.16. Against gpt-5.4 itself, the closed flagship, the gap stays wide and honest — 8.98 vs. 8.00 on Japanese MT-Bench, 4.41 vs. 3.79 on safety. A national research model beating an American lab's own open release on Japanese-language tasks, while trailing that same lab's closed frontier model by a clear margin, is a specific and legible result rather than a vague "competitive" claim.
What's still unstated
The card is explicit about being early-stage: "not tuned to ensure outputs align with human intent and safety considerations" beyond the disclosed benchmarks. There's no comparison against other sovereign or Japanese-market models — not Namazu, not Sakana's other releases, not Rakuten's or other domestic efforts. The model's position relative to Japan's own AI ecosystem, rather than against OpenAI's models, is still an open question. And the chat template's Harmony-format compatibility comes with a real caveat: the tokenizer differs from what the openai-harmony library expects, so direct tokenization with that library isn't supported. The companion cookbook repository makes the practical shape of that gap concrete. Running these models means either enabling trust_remote_code — the models bundle custom tokenizer and Harmony-parsing plugins to work correctly at all — or manually integrating that same code from the cookbook and the separate llm-jp-tokenizer project. That's a real deployment friction point for anyone in a trust_remote_code-averse environment, and it's not mentioned on the model card itself.
What to expect next
- Watch for independent benchmark reproduction, particularly the gpt-oss-20b comparison — it's the most checkable claim on the card and the one most relevant to anyone choosing between open Japanese-language models.
- Watch whether the no-RL approach gets revisited in a future generation. If NII adds an RL stage to LLM-jp-5 and reasoning scores jump, that's real evidence about what RL contributes; if they don't, that's evidence too.
- Watch how this sits alongside Namazu. Two different Japanese sovereignty strategies — train from scratch versus post-train someone else's weights — are now both live and both published with real numbers, a genuine natural comparison this blog can track over time.