Moonshot AI published the open weights for Kimi K3 on Hugging Face on July 27, eleven days after the model’s API debut. The release makes K3 the largest open-weight model publicly available to date, with 2.8 trillion total parameters and a 1 million-token context window. As we previewed yesterday, the drop is a milestone for China’s open-source AI push. But the details of what Moonshot actually shipped are more nuanced than the headlines suggest.
The architecture: 104B active parameters, not 50B
K3 is a sparsely activated Mixture-of-Experts model with 93 layers and 896 experts, with 16 experts selected per token plus two shared experts. The most important number that has been widely misreported: K3 activates 104 billion parameters per token, not the ~50B figure that circulated widely after the API launch. That distinction, documented in an engineering teardown by Aleksandro Matejic at Devoriales, is critical for anyone sizing inference clusters or estimating throughput.
The model uses two attention mechanisms across the stack: Kimi Delta Attention (KDA) on 69 of 93 layers and Gated MLA on the remaining 24, designed to sustain the 1 million-token context window without exploding compute. A 401-million-parameter MoonViT-V2 vision encoder makes K3 natively multimodal, suited for document-heavy and web-grounded workflows.
K3’s most unusual systems choice is its quantization regime: MXFP4 weights with MXFP8 activations, trained in from the supervised fine-tuning stage rather than applied post-training. That co-design shrinks the released weights to roughly 1.4 TB, small for a 2.8T-parameter model, while keeping the runtime within high-bandwidth memory budgets typical of large MoE inference. The tradeoff: Moonshot recommends 64 or more accelerators to serve the model, pushing true self-hosting into the realm of hyperscalers and well-funded labs.
Independent benchmarks: #3 overall, #1 on automation, higher hallucinations
On Artificial Analysis‘s independent evaluation, K3 scored 57 on the Intelligence Index, ranking third behind Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol. It reached 1668 Elo on GDPval-AA v2 and surged to 1547 Elo on AA-Briefcase, up 732 points over K2.6. On AutomationBench-AA, K3 topped the leaderboard at 53%, reinforcing Moonshot’s emphasis on tool use and long-context orchestration.
Efficiency gains are notable: K3 used 132 million output tokens to run the full index, 21% fewer than K2.6’s 166 million, and achieved the lowest cost per task among frontier peers at $0.94, compared to $1.04 for GPT-5.6 Sol and $1.80 for Opus 4.8. But there is a significant regression: a measured hallucination rate of 51%, up from K2.6’s 39%. For enterprise deployments in legal, medical, or compliance-heavy workflows, that uptick will shift more work to retrieval and verification layers.
Moonshot’s own benchmarks show a model that excels in specific areas: BrowseComp 91.2 (leads all), MCPMark-Verified 94.5 (leads all), SWE-Marathon 42.0 (leads all), and Harvey Lab-AA 94.6 (leads all). But K3 trails Fable 5 on FrontierSWE (81.2 vs 86.6), HLE-Full (43.5 vs 53.3), and even on Moonshot’s own Kimi Code Bench 2.0 (72.9 vs 76.9). The pattern is a model optimized for deep, tool-rich, agentic tasks rather than uniform dominance across abstract reasoning.
License: not MIT, and commercial use is gated
Despite being positioned as “open weights,” K3 does not use a permissive open-source license. The Kimi K3 licence allows commercial use for companies with under $20 million in annual revenue or under 100 million monthly active users. Above those thresholds, including Model-as-a-Service providers, a separate commercial agreement with Moonshot is required. Several outlets misreported the license as a modified MIT; the actual terms restrict frictionless adoption by larger platforms and could segment the K3 ecosystem between scrappy self-hosters and certified enterprise partners.
API pricing also jumped significantly: $3.00 per million input tokens and $15.00 per million output tokens, compared with K2.6’s $0.95/$4.00, a 3x to 3.75x increase. Moonshot launched K3 with a single “maximum thinking” effort level, which helps top-line quality but makes trivial prompts expensive by default.
What this means for China’s open-weight trajectory
K3’s release puts a China-based lab at the center of the global open-weight conversation on scale, on systems design, and on commercialization strategy. Founder Yang Zhilin‘s bet is that top-tier intelligence can be delivered in a more open form without fully open-sourcing the model. Whether that bet pays off depends on developer adoption, enterprise trust, and whether the restricted license deters the ecosystem effects that made DeepSeek’s open releases so consequential.
K3 is both a technical achievement and a policy statement. It raises the ceiling on what “open weights” can mean while redefining the floor on who can realistically run them. For China’s AI industry, it is a milestone that will force competitors and customers to make sharper choices about capability, cost, and control. The Kimi K3 API launch earlier this month provided the first look at what the model could do; the weights release now puts the question of who can actually run it front and center.