Moonshot AI releases the Kimi k1.5 reasoning model

A young company, a new target

Moonshot AI had been founded only in March 2023, less than two years before it shipped Kimi k1.5.^[1] By the time the model appeared, the centre of gravity in large language model work had moved from systems that answer straight away to systems that think for longer before answering: generating a long chain of intermediate reasoning steps and only then committing to a final answer. OpenAI's o1 was the reference point for this class of model, and matching it became an explicit goal for several labs.

The release

On 20 January 2025 Moonshot AI released Kimi k1.5.^[1][2] It is a multimodal reasoning model: it accepts images as well as text, and reasons over both.^[2] Moonshot claimed the model reached parity with OpenAI's o1 in mathematics, coding and multimodal reasoning.^[1][2] That was a competitive claim rather than a neutral benchmark report, but it set the terms on which the model was received.

The training recipe

Two days later, on 22 January 2025, the accompanying technical paper appeared on arXiv as "Kimi k1.5: Scaling Reinforcement Learning with LLMs".^[3]

The most distinctive part of the report is what it leaves out. Moonshot described reinforcement-learning methods that avoid Monte Carlo tree search, value functions and process reward models.^[1]

Those three components are the standard machinery of search-based reasoning training, and each does a different job. Monte Carlo tree search starts from a partial solution, explores many possible continuations, and keeps the branches that look promising. A value function is a learned estimate of how good an intermediate state is, and it is what guides that exploration. A process reward model scores each intermediate reasoning step rather than only the final answer, so the model receives feedback on where a chain of thought went wrong. All three cost extra training or extra computation: search has to run many branches during generation, while value functions and process reward models have to be trained separately. A recipe that works without them needs neither a separate search procedure at training time nor a separately trained step-level scorer.

The numbers

The paper's abstract reports two configurations, distinguished by how long a chain of thought the model is allowed to produce. The long-CoT version scored 77.5 on AIME and 96.2 on MATH 500. The short-CoT version, which reasons far more briefly, scored 60.8 on AIME and 94.6 on MATH 500.^[3]

The two benchmarks do not measure the same thing. MATH 500 and AIME are both collections of competition mathematics problems, and AIME is the harder and more selective of the two. The gap between the two configurations is therefore informative: on AIME the long chain of thought adds almost 17 points over the short one, while on MATH 500 the difference is only 1.6 points. Longer reasoning buys the most where the problems are hardest, and much less where a shorter solution already reaches the answer.

Why it matters

Two things stood out. The first is timing. A company founded in March 2023 was claiming parity with o1 at the start of 2025.^[1] Frontier reasoning capability was not confined to the largest and oldest laboratories, and the distance between a well-funded startup and the leading proprietary model was being measured in months.

The second is method. By setting out a reinforcement-learning approach that scales without tree search, value functions or process reward models, the report offered an alternative design point for anyone training a reasoning model.^[1] It also lent support to a broader argument: that much of the gain comes from the long chain of thought itself, with the elaborate search scaffolding around it optional rather than essential. The two configurations reported in the paper, and the size of the gap between them on hard problems, are the evidence for that argument.