Kimi k1.5

What Kimi k1.5 is

Kimi k1.5 is a reasoning model released by Moonshot AI on 20 January 2025 ^[1][2]. It is multimodal: it accepts images as well as text, and its reasoning is not confined to language inputs ^[2]. At release, Moonshot AI said the model reached parity with OpenAI's o1 in mathematics, coding and multimodal reasoning ^[1][2].

The model is documented in a technical report titled "Kimi k1.5: Scaling Reinforcement Learning with LLMs", credited to the Kimi Team and 95 other authors ^[3].

How it was trained

The report's title names the central idea: rather than leaning mainly on more pre-training data, the work scales reinforcement learning on top of a large language model ^[3]. Reinforcement learning, in this setting, means rewarding the outputs that score well on tasks a machine can check — a math answer that verifies, code that passes tests — and letting the model's parameters shift toward whatever produced the reward.

The authors describe a reinforcement-learning framework that works without three components commonly used in this kind of training: Monte Carlo tree search, value functions and process reward models ^[1][3]. In plain terms, Monte Carlo tree search explores many candidate branches of a reasoning path and chooses among them by simulating outcomes; a value function estimates how much future reward remains from a partly finished solution; a process reward model is a separately trained scorer that grades intermediate reasoning steps so the model can be rewarded step by step. Dropping all three simplifies the pipeline and removes the need to first train a step-level reward model — the reward comes from the final, checkable answer.

Reported results

The arXiv abstract reports two configurations ^[3]. The long chain-of-thought version, which spends many more tokens reasoning before it answers, scores 77.5 on AIME, 96.2 on MATH 500, reaches the 94th percentile on Codeforces and scores 74.9 on MathVista; the authors say these figures match OpenAI's o1 ^[3]. The short chain-of-thought version, which reasons more briefly, reports 60.8 on AIME, 94.6 on MATH 500 and 47.3 on LiveCodeBench ^[3]. The authors further claim that the short-CoT results outperform GPT-4o and Claude Sonnet 3.5 by up to 550% ^[3].

These are the authors' own reported numbers and comparisons. Benchmark scores depend on prompting, sampling and evaluation settings, so they describe the model as measured in that report.

Where it sits

The release placed Moonshot AI among the labs shipping models built around extended reasoning at inference time, and it pitched the work directly against OpenAI's o1 ^[1][2][3]. The technical report's contribution is as much methodological as numerical: it argues that strong reasoning can be trained with a comparatively simple reinforcement-learning loop, without search at training time and without a separately trained step scorer ^[1][3].

The k1.5 release was followed in April 2025 by the open-source Kimi-VL and in July 2025 by the open-weight Kimi K2 ^[2], as the Kimi family expanded into openly released multimodal and large models.