ELYZA-Thinking-1.0-Qwen-32B

ELYZA-Thinking-1.0-Qwen-32B is Japan's first specialized reasoning model, developed by ELYZA. It uses a Chain-of-Thought (CoT) approach, similar to OpenAI's o1 series, to tackle complex logical and mathematical problems. The model is built upon Alibaba's Qwen2.5-32B-Instruct and was fine-tuned on approximately 150,000 high-quality synthetic datasets generated using an innovative Monte Carlo Tree Search (MCTS)-based algorithm for optimal reasoning path exploration. This process enables the 32-billion parameter model to achieve performance comparable to OpenAI's o1-mini on key reasoning benchmarks, while remaining open-weight under the permissive Apache 2.0 license.

A key innovation is the dual-model approach: alongside the primary reasoning model, ELYZA released "Shortcut Models" (32B and 7B variants) trained on the same problem sets but without the lengthy reasoning process. The Shortcut model achieves performance comparable to GPT-4o on general tasks, demonstrating how complex reasoning capabilities developed during training can be distilled into faster, direct-response models. This work highlights a growing trend in AI development where heavy computational costs are shifted from inference to the development phase to create powerful yet efficient models.

While excelling in mathematical and logical reasoning, the model shows a trade-off: its coding capabilities slightly regressed compared to its base model, indicating the specialized training data may have lacked sufficient coding tasks. Nevertheless, it represents a significant milestone for Japanese-language AI, providing a powerful, commercially-viable open-source reasoning model that advances the state of the art for specialized inference tasks.