Moonshot AI released the full weights of Kimi K3 for public download today . It is the largest open-weight model ever released — 2.8 trillion parameters, 1 million token context window, and a front-end coding score that beat Claude Fable 5 .
The model that scored 57 on Artificial Analysis — third globally behind Fable 5 and GPT-5.6 Sol, ahead of Claude Opus 4.8 — is now free to download, modify, and deploy .
Anyone with sufficient infrastructure can now run a near-frontier model on their own servers. No per-token fees. No usage limits. No API gatekeeper.
The "Kimi Moment" Is Here
The release caps a two-week period that has fundamentally shifted the AI industry . On July 17, Kimi K3 launched and immediately topped Arena's front-end coding leaderboard with 1,679 points — ahead of Fable 5's 1,631 and GPT-5.6 Sol's 1,618 . Kimi K3's average task cost is $0.94, 65% cheaper than Fable 5 .
By July 22, nearly 200 US startups had signed a letter warning the White House that banning Chinese open models would kill "hundreds of companies" . On July 24, Jensen Huang posted his first X thread — not about GPUs, but to defend open-source AI .
The "Kimi Moment" is not just about Moonshot AI. It is about what a 2.8T open model does to the economics of the entire AI industry.
The Four Dimensions of the Shift
Performance: Kimi K3 leads all open models and rivals the best closed ones . It ranks first in front-end coding, and matches or exceeds Fable 5 on long-horizon coding and knowledge work benchmarks .
Cost: Each Kimi K3 task costs $0.94, vs $2.75 for Fable 5 and $1.04 for GPT-5.6 Sol . Its API pricing is higher than other Chinese open models ($3/$15 per million tokens), but Moonshot AI is betting its efficiency and performance justify the premium .
Architecture: Kimi K3 uses Kimi Delta Attention (KDA) and Attention Residuals, achieving 6.3x faster decoding at 1M context . The MoE design activates 16 of 896 experts per token, yielding 2.5x better scaling efficiency .
Openness: Moonshot AI's CEO Yang Zhilin has said the company's strategy is to win users through openness and availability . The full weights are now public. Enterprises can self-host and fine-tune freely .
The Debate Has Changed
OpenAI and Anthropic now face a different question than they did two weeks ago. It is no longer "should open models be regulated?" It is "what is our value proposition when a 2.8T open model matches our performance and costs 65% less?"
For the closed labs, the weight release is the worst-case scenario. For the open-source ecosystem, it is validation. For enterprise buyers, it is a new option.
P.S. If you are an AI developer, you can now run a model that beats Fable 5 on front-end coding. If you are an enterprise architect, you can now evaluate a near-frontier model without API costs. If you are a policy maker, you now have to decide whether to regulate a technology that anyone can download. The weight release is not the end of the debate. It is the beginning of a new one.
Frequently Asked Questions
Q: When can I download Kimi K3's weights?
A: Kimi K3's full weights are available for public download starting today, July 27, 2026, on platforms like Hugging Face. The model uses a Modified MIT license with minimal commercial restrictions.
Q: What are Kimi K3's specs?
A: 2.8 trillion total parameters, MoE architecture (896 experts, 16 activated per token), 1 million token context window, and native vision understanding.
Q: What hardware do I need to run Kimi K3?
A: You need a multi-GPU cluster. Even at MXFP4 quantization, the weights are around 1.4TB — a single H100 (80GB) is not enough. Moonshot AI recommends at least 64 accelerators per supernode.
Q: How does Kimi K3 compare to Fable 5 and GPT-5.6?
A: K3 scores 57 on the Artificial Analysis Intelligence Index — third globally behind Fable 5 (60) and GPT-5.6 Sol (59), ahead of Claude Opus 4.8 (56). On Arena's front-end coding leaderboard, it ranks first at 1,679 points, beating Fable 5's 1,631.
Q: What's the API pricing?
A: $3 per million input tokens, $15 per million output tokens. Cache-hit input is $0.30 per million. Single-task average cost is $0.94 — about one-third of Fable 5's $2.75.
Q: What are K3's known limitations?
A: Moonshot AI acknowledges K3 still trails Fable 5 and GPT-5.6 on overall performance. Independent testing shows hallucination rates around 51%. K3 also tends to over-optimize — consuming excessive tokens on simple tasks and occasionally expanding task scope beyond what was requested.
Q: How does K3 compare to Fable 5 in practice?
A: K3 matches or exceeds Fable 5 on front-end coding and knowledge work (AA-Briefcase: 1,543, second only to Fable 5), but still lags on complex reasoning and long-horizon agent tasks. Generation speed is also noticeably slower.
Q: Is Microsoft testing Kimi K3?
A: Yes, reports indicate Microsoft is internally testing Kimi K3 and evaluating whether to integrate it into Copilot AI assistant to reduce inference costs.
