ELYZA free-releases ELYZA-Thinking-1.0-llm-jp-4-33b models

ELYZA built these models by taking the LLM-jp-4 series—an open model developed by Japan's National Institute of Informatics—and running additional intermediate and post-training focused mainly on Japanese-language processing. That approach matters because it lets a company specialize an existing open base rather than train from scratch.

The claim that the two models outperform NII's own improved version, LLM-jp-4.1, on some benchmarks is notable because LLM-jp-4.1 was only announced on September 28, just days before. It suggests the gap between a research institute's base model and a corporate post-trained version can be narrow, though the benchmarks in question are not named here.

The release is also tied to ELYZA's new ELYZA RSI Research organization, which pursues recursive self-improvement. ELYZA describes that effort as still at an element-research stage, so readers should treat the RSI framing as a long-term direction rather than an achieved capability. The practical test will be how the two models perform on real Japanese tasks, not just on the cited benchmarks.