Moonshot AI releases Kimi K1.5 reasoning model

Moonshot AI, the Beijing lab behind the Kimi chatbot, published Kimi k1.5, a reasoning model the company said matched OpenAI’s o1 across multiple modalities on mathematics, coding and visual reasoning benchmarks including AIME, MATH-500, LiveCodeBench and MathVista. It shipped the same day DeepSeek released R1, a coincidence of timing that meant Kimi k1.5 received comparatively little attention despite overlapping claims.

The accompanying technical report described a reinforcement-learning recipe Moonshot characterised as simpler than competing approaches: no Monte Carlo tree search and no separately trained value function, context scaled to 128,000 tokens using “partial rollouts” that reused earlier reasoning trajectories to cut training cost, and online mirror descent for policy optimisation with length penalties to control response size. The model was trained jointly on text and vision data, and Moonshot reported large margins over GPT-4o and Claude 3.5 Sonnet on shorter chain-of-thought responses, citing a gain of up to 550% on one benchmark comparison.

Kimi k1.5 was not released with open weights — Moonshot published the method and results rather than a downloadable model, which set it apart from the fully open release DeepSeek made the same week. It was one of several reasoning models to come out of Chinese labs within days of each other that January, alongside DeepSeek-R1 and, a week later, Alibaba’s Qwen2.5-Max, as competition over reasoning benchmarks intensified.