Japan's sovereign LLM project LLM-jp releases open 4.1 models with tool calling
markjeffrey · x · 2026-09-29
Japan's sovereign LLM effort LLM-jp, led by NII's LLMC and Prof. Sadao Kurohashi, released the open-source LLM-jp-4.1 series under Apache 2.0.
- Three thinking models: 8B Dense, 32B-A3B MoE, and 33B Dense, all on Hugging Face.
- Key upgrades: improved STEM performance (math, science, instruction-following beats LLM-jp-4 at same scale), new tool-calling support for agent systems, and more concise answers.
- SFT and DPO training datasets are also published; API and chat access will roll out via the University of Tokyo's mdx-MaaS from mid-October.
- Next-gen LLM-jp-4.2 plans to add reinforcement learning for better reasoning and agentic capabilities.
More from Models
- Frontier models refuse to harden Windows DCs 43.8% of the time, more if you claim authorization — Aizkmusic · 2026-09-29
- OpenAI DevDay agenda leaks: Codex to get platform capabilities for plugins, agents and apps — testingcatalog · 2026-09-29
- OpenAI discloses GPT-6.1 Astra showed deception and overstepped user authorization — soumitrashukla9 · 2026-09-29
- Same Prompt, 21 Replays: Sonnet 5.5 Took the Closest Pokémon 13 Times, Opus 5.5 Never Did — VibeCodyH · 2026-09-29
- ISTA-DASLab's pruned, quantized Qwen3.8 Coder GGUF model trends on Hugging Face — ISTA-DASLab · 2026-09-29
- Anthropic's real moat may be alignment that doesn't lobotomize the model — 'Teaching Claude Why' cut misalignment 19x — PrisonOfH0pe · 2026-09-29