Moonshot AI launched its flagship Kimi K3 model on July 16, 2026. The model has 2.8 trillion total parameters, native vision support, and a 1-million-token context window. Kimi K3 is already available through Kimi, Kimi Work, Kimi Code, and the official API. At launch, the full model weights were still being prepared and were scheduled for public release by July 27.
How Does Kimi K3's 2.8-Trillion-Parameter Architecture Work?
Kimi K3 uses a Mixture of Experts (MoE) architecture that activates 16 of 896 experts with Stable LatentMoE. Its attention stack combines Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). KDA provides an efficient basis for long-sequence attention, while AttnRes selectively retrieves representations from different network depths.
Moonshot AI says these components, together with updated training and data recipes, deliver an approximately 2.5x improvement in overall scaling efficiency over Kimi K2. Quantization-aware training begins at the SFT stage with MXFP4 weights and MXFP8 activations. The company recommends supernode deployments with at least 64 accelerators, putting full-weight inference mainly within reach of large inference providers and research organizations.
How Does Kimi K3 Perform on Coding and Agent Benchmarks?
In Moonshot AI's published evaluation, Kimi K3 (max) scored 91.2 on BrowseComp, 88.3 on Terminal-Bench 2.1, 81.2 on FrontierSWE, and 42.0 on SWE Marathon. Long-horizon coding demonstrations included GPU kernel optimization, building the MiniTriton compiler from scratch, and a 48-hour autonomous run that designed, optimized, and verified a 45nm chip for a nano model.
The numbers require careful interpretation. Moonshot AI acknowledges that Kimi K3 still trails Claude Fable 5 and GPT 5.6 Sol overall. Some coding evaluations also used different harnesses across models, including Kimi Code, Claude Code, and Codex. The vendor-reported results place K3 in the frontier-model discussion, while independent and reproducible third-party testing is still needed.
How Can Developers Access Kimi K3 Through the API?
The official API model ID is kimi-k3, and the service is compatible with the OpenAI API format. Reasoning is always enabled in the launch version, with reasoning_effort currently limited to max. Multi-turn conversations and tool calls must return the complete assistant message to the model; dropping its reasoning history can cause a marked decline in output quality.
Global pricing per million tokens is $0.30 for cache-hit input, $3 for cache-miss input, and $15 for output. Moonshot AI says its Mooncake disaggregated inference architecture produces a cache-hit rate above 90% on coding workloads. The Chinese API documentation also warns that the web-search tool is being upgraded and is not currently recommended for production use.
What Limitations Did Moonshot AI Disclose at Launch?
Moonshot AI lists three limitations with direct consequences for Agent deployments:
- Sensitivity to thinking history: The Agent harness must preserve and return the full reasoning history. Switching to K3 from another model in the middle of a session may destabilize output.
- Excessive proactiveness: K3 may make unexpected decisions when it encounters small issues or ambiguous instructions. Applications need explicit operating boundaries in the system prompt or AGENTS.md.
- Remaining experience gap: Moonshot AI says K3 still has a noticeable user-experience gap compared with Claude Fable 5 and GPT 5.6 Sol.
Kimi K3 pushes announced open-weight model scale close to the 3T mark while combining long context, vision, and long-horizon Agent work in one system. The hosted products and API can be evaluated immediately. Claims about openness, deployment cost, and real-world stability will remain provisional until the full weights, technical report, and independent evaluations are available.