A chatbot that can solve competition math and write working code is no longer the headline. The interesting question is how cheaply a lab can get there, and what tricks it had to use to do it.
Moonshot AI (月之暗面), the Chinese lab behind Kimi, published a model in early 2025 that sits in the same reasoning tier as OpenAI's o1.
What makes it worth a closer look is not the scoreboard but the recipe: the team says it reached that tier without three techniques almost everyone assumed were necessary.
Key takeaways
- Kimi k1.5 is a multimodal large model (大模型) released by Moonshot (月之暗面) on 20 January 2025, trained with reinforcement learning (强化学习).
- On its long-chain-of-thought setting it posts 77.5 on AIME 2024, 96.2 on MATH-500 and a 94th-percentile Codeforces rating—numbers the company's technical report says match OpenAI's o1.
- A shorter, faster setting (60.8 AIME, 94.6 MATH-500) beats GPT-4o and Claude Sonnet 3.5 by as much as +550% on the same benchmarks.
- The method scales the RL context window to 128k tokens and uses a simpler policy-optimization loop that, the team says, needs no Monte Carlo tree search, value functions, or process reward models.
- Moonshot shared the full training recipe in an open technical report (arXiv:2501.12599), making it one of the few o1-class efforts with a published method.
- It was deployed through the Kimi product and later distilled variants, not released as full open weights.
A reasoning model that reads images
Kimi k1.5 is described by its makers as a multi-modal LLM trained on both text and vision data, which means it can reason over images as well as text. That placed it, at launch, among a small group of reasoning models that could take image inputs—a capability the contemporaneous o1 preview treated as text-only. The paper reports 74.9 on MathVista and 70.0 on MMMU, multimodal benchmarks where it trailed o1's larger general-knowledge mix but still landed in the same league.
This matters because most real reasoning workloads—reading a chart, checking a screenshot, interpreting a diagram—start with an image, not a paragraph. A model that can reason across both modalities is closer to how people actually work.
The recipe: scale context, simplify the loop
The technical report names two ingredients. First, long-context scaling: Moonshot pushed the RL training context window to 128k tokens and reported continued improvement as length grew. The trick that keeps this affordable is "partial rollouts"—reusing a large chunk of a previous trajectory instead of regenerating it, which cuts the cost of longer reasoning traces.
Second, improved policy optimization. The team derives an RL formulation for long chains of thought and uses a variant of online mirror descent, plus sampling strategy, length penalty, and data-recipe tuning. The claim that matters: strong results are achievable without Monte Carlo tree search, value functions, or process reward models—the heavier machinery many labs assumed was required. Whether that holds outside Moonshot's own benchmarks is exactly what independent groups are now testing.
Why "long2short" is the practical part
The headline o1-matching scores come from the long-chain setting, which is slow and expensive at inference. The more useful contribution may be the "long2short" methods: using the long-chain model to generate compact reasoning traces, then fine-tuning a short-chain variant on them. The result—60.8 AIME, 94.6 MATH-500, 47.3 LiveCodeBench—runs far cheaper while still beating GPT-4o and Claude Sonnet 3.5 by up to 550% on those tasks. The short-chain variant is the default in the consumer Kimi app for routine queries, which is the version most users actually meet.
The model itself was deployed primarily through the Kimi product and later through distilled variants on Hugging Face, rather than as a full open-weight base model. For builders who want to experiment, that means the distillation path, not a drop-in weights download.
Honest limitations
This article leans on Moonshot's own technical report (arXiv:2501.12599) and launch coverage from People's Daily and Xinhua Finance for dates and the "second lab after OpenAI to reach this level" framing. The benchmark numbers are company-reported; external reproductions have matched the ranking but vary slightly with decoding settings and prompts. We have not re-run the benchmarks. The comparison to o1 uses OpenAI's published figures from late 2024 and early 2025, a fast-moving baseline. Capability gaps versus the newest closed models change monthly, so treat the specifics as a marker of a trend, not a current scoreboard.
What readers can do now
- If you build with reasoning models, test Kimi k1.5's short-chain setting via the Kimi OpenPlatform before defaulting to a premium API—the cost per solved problem is often an order of magnitude lower.
- When a lab claims "o1-level" performance, check which benchmark and which setting; long-chain and short-chain scores differ by a wide margin.
- For multimodal tasks (charts, screenshots, diagrams), prototype a vision-capable reasoning model directly instead of preprocessing images into text.