Key Points
- Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter Mixture-of-Experts model
- Full open weights are due by July 27 under a Modified MIT license
- K3 ranks first on Arena.ai's Frontend Code leaderboard with 1,679 points
- Input tokens cost $3 per million on a cache miss; output tokens $15
Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter Mixture-of-Experts model that the Beijing company shipped simultaneously across its consumer apps and its developer API. Tom's Hardware and VentureBeat both described it as the largest open-weight model released to date. The weights themselves are not out yet: Moonshot has committed to publishing them by July 27.
The architecture
K3 is a sparse Mixture-of-Experts model built on what Moonshot calls its Stable LatentMoE framework: 2.8 trillion total parameters spread across 896 experts, of which 16 activate on any given token, roughly 1.8% of the pool. That is the design that lets a model of this size run at a serviceable cost — the parameter count describes what the model knows, not what it computes per token.
Two in-house components sit underneath it. Kimi Delta Attention (KDA) is a hybrid linear attention mechanism, and Attention Residuals (AttnRes) is a replacement for standard residual connections; Moonshot says both are aimed at how information moves across sequence length and model depth. The context window is 1 million tokens. The model takes text, image and video input, and it runs at maximum thinking effort by default, with low- and high-effort modes slated for later updates. A full technical report on architecture, training and evaluations is due alongside the weights.
What the benchmark placements measure
Arena.ai put K3 at No. 1 on its Frontend Code leaderboard with 1,679 points, ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618). That board ranks models by blind head-to-head human preference on front-end and web-building prompts; Arena also reported a 76% pairwise win rate for K3 against 63% for Fable 5 and 58% for GPT-5.6 Sol, a first-place finish in six of seven domains, and second only in Gaming. It is a 17-place jump from Kimi K2.6, which sat at No. 18.
The broader picture is more measured. On the Artificial Analysis Intelligence Index — a composite of nine evaluations spanning agentic tasks and knowledge work — K3 scored 57, placing third behind Claude Fable 5 and GPT-5.6 Sol and level with Claude Opus 4.8 and GPT-5.5. Artificial Analysis measured an average cost of $0.94 per task and found K3 used 21% fewer output tokens than K2.6 across the index, about 132 million versus 166 million. Writing in Interconnects, Nathan Lambert placed K3 second overall on the Vals AI index.
Moonshot's own claim is narrower than the headlines. Its benchmark comparisons, reported by CNBC, put K3 behind Claude Fable 5 and GPT-5.6 Sol on overall performance, while beating Claude Opus 4.8 and GPT-5.5 on benchmarks including coding and general agents.
Pricing and access
The API runs $3 per million input tokens on a cache miss, $0.30 per million on a cache hit — a 90% discount — and $15 per million output tokens. For comparison, Anthropic launched Claude Fable 5 in June at $10 per million input and $50 per million output, which puts K3 at roughly a third of the rate of the model it beat on Arena's front-end board.
K3 is live now on kimi.com, Kimi Work, Kimi Code and the Kimi API under the model ID kimi-k3, with apps on iOS, Android and HarmonyOS and desktop builds for Windows and Apple silicon Macs.
The license and the weight drop
Moonshot has said the weights will ship by July 27 under a Modified MIT license, per Notebookcheck. Until that date, K3 is a frontier model you can rent but not download. The release lands three weeks after Washington moved the other way on American models, asking OpenAI to limit GPT-5.6's initial rollout to about 20 government-approved partners.
Who Moonshot AI is
Moonshot AI is a Beijing startup founded in 2023, backed by Alibaba, Tencent and Meituan. Forbes reported the company was valued above $20 billion in a May round and that its annual recurring revenue passed $200 million in April 2026; by June it was in talks to raise up to $2 billion at a $30 billion valuation. Bank of America analysts led by Alex Liu wrote in a note cited by CNBC that despite "persistent hardware/compute capacity constraints in China, K3 demonstrates that pre-training scaling, paired with architectural innovation, can still deliver step-change gains for flagship Chinese models."
Others covering this
CNBC — "China's Moonshot AI unveils Kimi K3 that rivals OpenAI, Anthropic"
"It's the latest AI model from China to close the performance gap with leading U.S. AI labs."
The New Stack — "Kimi K3 tops Arena's coding leaderboard — and it's open-weight"
"Open-weight model Kimi K3 tops Arena's frontend coding leaderboard, challenging proprietary AI coding tools and pushing IDE vendors to rethink model lock-in."
Notebookcheck — "Kimi K3 tops Frontend Code Arena in a first for Chinese AI models"
"Moonshot AI's Kimi K3 has climbed into the top 10 frontier text rankings, rivaling Claude, ChatGPT, and Gemini. It's also shot to the top of the frontend coding leaderboard, becoming…"
Interconnects — "Kimi K3: The open-weights escalation"
"The global implications on the AI ecosystem."
The open-weights fight escalated within a week: Nvidia launched the Open Secure AI Alliance on July 27 with 74 inaugural partners and without OpenAI, Google or Anthropic.
Source: Moonshot AI, Artificial Analysis, Arena.ai.
Comments