Kimi K3 Benchmarks: A Step Forward for Chinese Frontier AI
Chinese AI labs have spent the past year narrowing the gap with the best proprietary models. DeepSeek, Qwen, GLM, MiniMax, and Kimi have each been competitive in different areas.
Kimi K3 is the clearest step so far.
Moonshot AI's new flagship is a 2.8-trillion-parameter model built for long coding sessions, tool use, research, visual work, and other tasks that may take many steps to finish. In Moonshot's benchmarks, it is competitive with leading proprietary models and often well ahead of GLM-5.2, one of the strongest recent models from a Chinese lab.
The numbers justify the attention, but not every launch-day headline.
What changed with Kimi K3
Kimi K3 is a mixture-of-experts model. It has 2.8 trillion parameters in total, but it does not use all of them for every token. Moonshot says it activates 16 of 896 experts at a time, allowing it to reach a very large overall scale without paying the full computational cost on every step.
For most users, the practical changes matter more than the architecture:
- A 1-million-token context window for large repositories, long documents, and extended work
- Native understanding of text, images, and video, according to Moonshot; NanoGPT accepts all three as inputs
- Reasoning, tool calling, and structured outputs on NanoGPT
- A stronger focus on long-running work rather than one-shot answers
- Max reasoning effort, currently the only K3 reasoning level available on NanoGPT
Moonshot describes K3 as the first open model in the 3-trillion-parameter class. There is an important timing detail, though: Moonshot says the full weights will be released by July 27, 2026. As of this article's publication, the hosted model is available but the weights are not yet downloadable. It is more accurate to call this an announced open-weight release until that happens.
Independent testing puts it near the top
Artificial Analysis currently gives Kimi K3 a score of 57 on its Intelligence Index, placing it fourth among the 189 models shown on the model page when we checked on July 17. Rankings can change as new models and results are added, but that is still a strong independent signal that K3 belongs near the current frontier.
The same testing also highlights the tradeoffs:
- K3 produced about 62 output tokens per second in the tested first-party API, below the 71.1-token comparison median shown on the page.
- It was unusually verbose, generating 130 million evaluation output tokens versus a comparison median of 63 million.
- Artificial Analysis lists both its input and output pricing above the median for comparable models.
In other words, K3 is not the obvious choice for every prompt. Its case is capability, not minimal cost or instant answers. The independent result is also a more useful starting point than treating Moonshot's launch charts as neutral evidence.
Moonshot reports a large coding and agent jump
Moonshot published a broad set of coding, agent, knowledge-work, and visual benchmarks. The clearest comparison for the progress of models from Chinese labs is GLM-5.2, which appears throughout the same tables.
| Benchmark | Kimi K3 | GLM-5.2 | What it measures |
|---|---|---|---|
| Terminal-Bench 2.1 | 88.3 | 82.7 | Completing practical tasks in a terminal |
| FrontierSWE | 81.2 | 67.3 | Solving difficult real-world software issues |
| Program Bench | 77.8 | 63.7 | Writing and reasoning about programs |
| SWE Marathon | 42.0 | 13.0 | Sustained work on longer software tasks |
These are Moonshot's own reported results, with K3 running at maximum reasoning effort. They are useful, but they are not a perfectly controlled head-to-head test. Different models sometimes used different coding-agent harnesses, and the amount of scaffolding around a model can affect the result.
That caveat matters most for the largest gaps. A score of 42.0 against 13.0 on SWE Marathon is striking, but it should encourage independent testing rather than end the comparison.
Moonshot is fairly direct about the overall picture: K3 still trails Claude Fable 5 and GPT-5.6 Sol when everything is considered together. The interesting part is that it is now close enough to compete with those models on several demanding tasks and lead some of the published tables.
Why this matters for Chinese AI models
The strongest Chinese-developed models have often been discussed as impressive alternatives: capable enough for serious work, open or partly open, and usually cheaper than the leading Western proprietary models.
Kimi K3 changes the emphasis. It is trying to compete at the top on long-horizon coding, tool use, knowledge work, and multimodal tasks, not merely offer a better price-to-performance ratio.
The comparison with GLM-5.2 is particularly notable. GLM-5.2 is already a strong engineering and agent model. K3's reported gains across terminal work and several kinds of software tasks suggest a broader improvement rather than one benchmark-specific trick.
Parameter count alone does not make a model useful, and a promised open-weight release is not the same as weights people can inspect and run today. But if the weights arrive as announced and independent evaluations continue to support the launch results, K3 will be strong evidence that models from Chinese labs can compete near the frontier on demanding coding and agent tasks.
Where Kimi K3 makes sense
K3 is worth testing when the task can benefit from sustained reasoning and a lot of context:
- Working across a large codebase
- Debugging or implementing a feature over many steps
- Combining screenshots, video, documents, and code
- Running tool-heavy research or analysis
- Building an agent that needs to keep a long project in view
- Reviewing a large amount of material before producing a final result
It is less attractive for quick questions, short rewrites, simple extraction, or other tasks where a faster and cheaper model already works. Kimi K3 is also pay-as-you-go on NanoGPT and is not currently included in the subscription.
Two practical caveats
Moonshot lists two limitations that are worth taking seriously.
First, K3 expects its earlier reasoning history to remain available during a continuing task. Switching to it halfway through a conversation—or using a client that drops earlier reasoning state—can make its output unstable. Start a fresh K3 conversation for important work and avoid changing models in the middle of a long run.
Second, K3 can be overly proactive. It was trained to keep difficult tasks moving, which can also mean making decisions you did not explicitly ask it to make. Give it clear boundaries when it can edit files, run tools, spend money, publish work, or take any other consequential action.
Those are manageable limitations, but they reinforce the larger point: this is an agent-oriented model, not just a chatbot with a bigger context window.
The short version
Kimi K3 is one of the most convincing frontier models from a Chinese lab so far. Moonshot's benchmarks show a large step beyond GLM-5.2 on several coding and agent tasks, while independent testing places it near the top overall.
It is not yet a fully released open-weight model, it is not especially fast or concise, and it will be excessive for simple work. But for long coding sessions, complex tool use, multimodal projects, and other tasks where a model needs to keep going rather than answer once, it is worth evaluating against the other frontier options.
You can try Kimi K3 on NanoGPT now. API users can select moonshotai/kimi-k3 through NanoGPT's OpenAI-compatible API.
