Kimi K3

Kimi K3, launched July 14, 2026, is Moonshot AI's flagship open-weight model and the world's first open 3T-class model. With 2.8 trillion total parameters (104B active), native vision, and a 1M-token context window, it brought frontier-level agentic and coding performance to downloadable weights. Its full weights landed on Hugging Face on July 27, 2026, quickly becoming one of the most-downloaded open models of the year.

Model Specifications

Key Features

• Frontier open weights: A 2.8T model anyone can download, deploy, fine-tune, and build on; the first open model to cross the 3T-class threshold.
• Always-on thinking: Reasoning effort of low, high, or max (default), with no way to disable thinking entirely.
• Native multimodality: Text, images, and video in one model, with strong document and UI rendering performance.
• Efficient serving: MXFP4/MXFP8 quantized training plus a vLLM prefix-cache implementation for low-cost 1M-token inference.

Benchmark Highlights

• Terminal-Bench 2.1: 88.3, effectively tied with the best closed models.
• BrowseComp: 91.2 with context compaction, ahead of Claude Fable 5 (88.0).
• MCPMark-Verified: 94.5, the best agentic tool-use score in its comparison set.
• SWE-Marathon: 42.0, the top result among Fable 5, GPT-5.6 Sol, and GLM-5.2.
• Artificial Analysis Intelligence Index: 57, third behind Claude Opus 5 (61) and Fable 5 (60), ahead of Claude Opus 4.8 (56).

Pricing

Kimi K3 on the Kimi API is $3.00/MTok for cache-miss input, $0.30/MTok for cache-hit input, and $15.00/MTok for output, with automatic prefix caching. It is also included in Kimi's monthly subscription plans and available through Kimi.com, Kimi Work, and Kimi Code.

Licensing Caveat

The Kimi K3 License is permissive for researchers, startups, and internal use, but companies operating a Model-as-a-Service business with more than $20M annual revenue must sign a separate commercial agreement with Moonshot AI, and the largest deployments (100M+ MAU or $20M+ monthly revenue) must display "Kimi K3" prominently in their UI.

Getting Started

Download weights from the MoonshotAI/Kimi-K3 Hugging Face repository or use it via the Kimi API with reasoning_effort set to low, high, or max. For local inference, pair it with vLLM or SGLang (both supported) and expect roughly 1.5 TB of weights, which suits clusters or high-end workstations rather than a single consumer GPU.

Alternatives

• Claude Opus 5: Closed frontier model with comparable agentic performance at $5/$25 per MTok.
• GLM-5.2: Another leading open-weight Chinese model with strong coding scores.
• DeepSeek V4: Lighter open-weight option with a very permissive license.

Conclusion

Kimi K3 proved that open weights can sit within striking distance of the closed frontier on agentic and coding workloads. For teams with the infrastructure to run 1.5 TB of weights, it is the most capable open model available, at a fraction of the API cost of comparable closed models.