Kimi K3, launched by Moonshot AI on 16 July 2026, is the first open-weight model to rank #1 on a major human-preference coding leaderboard, taking the top slot on Arena.ai's Frontend Code Arena with 1,679 points — ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618). It is a 2.8T-parameter MoE model with a 1M-token context window and native vision. Moonshot has committed to releasing the full weights by 27 July 2026. That combination — frontier coding performance, permissive distribution, Sonnet-tier pricing — is what makes this release structurally different from earlier open-weight launches.
TL;DR
- Kimi K3 is the largest open-weight model ever: 2.8T total parameters, 16 of 896 experts active per token.
- Ranks #1 on Arena.ai's Frontend Code Arena at 1,679 Elo, ahead of Claude Fable 5, GPT-5.6 Sol, GLM-5.2, and Grok 4.5.
- Artificial Analysis places it at 1,547 Elo on long-horizon knowledge work, second only to Claude Fable 5.
- Full weights ship 27 July 2026; API is live at $3/M input and $15/M output tokens.
- Thinking mode is always on;
reasoning_effortonly acceptsmaxat launch. - Moonshot's own framing: K3 trails Fable 5 and GPT-5.6 Sol overall, but shows frontier-level performance in specific domains.
What is Kimi K3?
Kimi K3 is Moonshot AI's flagship model, released on 16 July 2026. It is a sparse MoE model with 2.8 trillion total parameters, of which 16 of 896 experts are activated per token via the Stable LatentMoE framework. The model introduces Kimi Delta Attention (KDA) for information flow across long sequences and Attention Residuals (AttnRes) for signal preservation at depth. Moonshot reports roughly 2.5× improvement in scaling efficiency vs Kimi K2.
Two variants ship at launch: K3 Max (chat/agent, via kimi.com and the OpenAI-compatible API) and K3 Swarm Max for large-scale parallel processing. Both support 1M-token context and native vision.
Why is Kimi K3 significant for the open-weight ecosystem?
Until now, the argument for open-weight models has been mostly economic. You could self-host, fine-tune, and avoid vendor lock-in, but you accepted a capability gap versus closed frontier models. Kimi K3 changes that. On Arena.ai's Frontend Code Arena — a public, human-preference coding leaderboard — an open-weight model is now in first place, ranked first in six of seven front-end domains.
It matters for three reasons:
- Capability parity, not just cost parity. Teams that ruled out open weights on quality grounds now have to re-evaluate.
- No vendor lock-in at frontier tier. Once weights land on 27 July 2026, organisations can run K3 on their own infrastructure or fine-tune it for internal codebases.
- Pricing signals confidence. At $3/$15 per million tokens, K3 is priced in Anthropic's Claude Sonnet tier. Moonshot is no longer discount-pricing its flagship.
The broader context is that the open-weight ecosystem is widening. Our free AI coding agents omniroute strategy covers how to combine open and closed models in a single agent stack.
How does Kimi K3 compare to Claude Fable 5 and GPT-5.6 Sol?
On the leaderboard that started the conversation, K3 beats both. On broader benchmarks, the picture is more mixed and worth reading carefully.
| Benchmark | Kimi K3 | Notes |
|---|---|---|
| Frontend Code Arena (Elo) | 1,679 | #1, ahead of Fable 5 at 1,631 and GPT-5.6 Sol at 1,618 |
| GPQA Diamond | 93.5% | Strongest open-weight score reported |
| Terminal-Bench 2.1 | 88.3% | Behind GPT-5.6 Sol (88.8%) |
| BrowseComp | 91.2% | Best published score |
| Humanity's Last Exam (with tools) | 56.0% | — |
| MCP Atlas | 84.2% | — |
| FrontierSWE | 81.2% | — |
| MMMU-Pro | 81.6% | — |
| Artificial Analysis Elo (long-horizon) | 1,547 | +732 from K2.6; second only to Claude Fable 5 |
The pattern: K3 is strongest on front-end coding, browsing, and knowledge-heavy benchmarks. Moonshot itself states the model "trails the most powerful proprietary models" overall while showing frontier-level performance in specific areas.
Artificial Analysis reports a per-task cost of $0.94 for K3, versus $1.04 for GPT-5.6 Sol and $1.80 for Opus 4.8. For a side-by-side of the closed models K3 is chasing, see GPT-5.6 Sol vs Claude Fable 5 business comparison and Fable 5 vs Grok 4.5 vs GPT-5.6 comparison guide.
What can Kimi K3 actually do in practice?
Two case studies from Moonshot test the model against real engineering. In GPU kernel optimisation, K3 matched Fable 5 (which needed fallback) and beat Opus 4.8, GPT-5.6 Sol, and GPT-5.5. An early K3 build handled most of Moonshot's own kernel optimisation during development.
K3 also built MiniTriton, a compact Triton-like GPU compiler from scratch with tile-level IR, optimisation passes, and PTX codegen — performing on par with or better than Triton and torch.compile.
If you are choosing where K3 fits in your editor workflow, our roundup of the best AI code editor 2026 covers the integration surface.
How much does Kimi K3 cost and how do you use it?
The API is live at kimi.com and through the OpenAI-compatible Kimi API:
- Input: $3/M tokens | Cached input: $0.30/M | Output: $15/M
That places K3 in the Claude Sonnet tier — the most expensive model a Chinese lab has priced to date, signalling Moonshot is positioning K3 as a frontier product, not a value alternative.
Two implementation details matter at launch: thinking mode is always enabled (reasoning traces stream as separate deltas), and reasoning_effort only accepts max.
For agent architectures that route across multiple frontier models, GPT-5.6 Sol vs Fable 5 agent strategy covers the decision framework you would extend to include K3.
When will Kimi K3 weights be released, and under what licence?
Moonshot has committed to releasing full weights by 27 July 2026. The licence is unconfirmed — K2 used Modified MIT, so that is the working assumption. Moonshot AI is Beijing-based, backed by Alibaba and Tencent, with a $500M Series C at $4.3B valuation in January 2026.
FAQ
Q: Is Kimi K3 really open-source? A: Moonshot has committed to releasing full weights by 27 July 2026. The licence is unconfirmed — K2 used Modified MIT. Wait for the licence text on release day before assuming redistribution rights.
Q: How does Kimi K3 rank against Claude Fable 5? A: On Arena.ai's Frontend Code Arena, K3 leads by 48 Elo (1,679 vs 1,631). On Artificial Analysis's long-horizon evaluation, Fable 5 leads. Moonshot states K3 trails Fable 5 and GPT-5.6 Sol overall but matches or beats them in specific domains.
Q: Can I self-host Kimi K3? A: Once weights land on 27 July 2026, yes — but a 2.8T MoE needs a substantial multi-GPU setup even in serving mode. Hosting providers are expected to offer managed inference shortly after.
Q: Why can't I turn off thinking mode?
A: Moonshot ships K3 with reasoning always enabled. The streaming API exposes reasoning traces as separate deltas. reasoning_effort accepts only max on day one — expect configurability to expand later.
Q: Is Kimi K3 cheaper than GPT-5.6 Sol or Claude Opus 4.8? A: At $3 input and $15 output per million tokens, K3 is in the Sonnet tier — cheaper than Opus 4.8, comparable to Sol. Artificial Analysis reports per-task cost of $0.94 for K3 vs $1.04 for Sol and $1.80 for Opus 4.8.
Q: What is Kimi Delta Attention? A: KDA improves information flow across long sequences; Attention Residuals (AttnRes) improve flow across model depth. Together they contribute to a reported 2.5× scaling efficiency gain over Kimi K2.
Discussion
0 comments