Kimi K3

Quick Answer

Kimi K3 is Moonshot AI’s open-weight frontier model, released July 17, 2026. At 2.8 trillion total parameters it is reported to be the largest open-weight model released to date, built with a hybrid linear attention design (Kimi Delta Attention plus Attention Residuals), native vision support, and a 1M-token context window. Moonshot’s own benchmarks place it behind only Claude Fable 5 and GPT-5.6 on general capability, and ahead of Claude Opus 4.8 and GPT-5.5 on coding and agent tasks.

It is a different thing from Kimi Code, Moonshot’s terminal coding agent. Kimi K3 is the underlying model; Kimi Code is the CLI tool that can run on Moonshot’s models, K3 included, through the same Moonshot AI Open Platform access.

What Kimi K3 Is Best For

• Open-weight frontier capability: near-top-tier benchmark results without a closed-provider dependency
• Coding and agentic workflows: strong reported results on coding and general-agent benchmarks
• Long-context and multimodal tasks: 1M-token context window with native visual understanding
• Reducing vendor lock-in: a credible open alternative for teams building model fallback plans

Key Facts (Verified)

• Developer: Moonshot AI
• Released: July 17, 2026, with full weight release scheduled to complete by July 27, 2026
• Type: Open-weight Mixture-of-Experts model, 2.8 trillion total parameters
• Architecture: Kimi Delta Attention (a hybrid linear attention mechanism) plus Attention Residuals
• Context window: 1M tokens, with native visual (vision) understanding
• Access: Kimi.com, the Moonshot AI Open Platform API (platform.kimi.ai), and downloadable weights once the release completes

Benchmark comparisons against Claude Fable 5, GPT-5.6, Claude Opus 4.8, and GPT-5.5 come from Moonshot’s own reporting. As with any vendor-published benchmark, verify against independent evaluations and your own workloads before making a switching decision.

How to Access It

Easiest: chat with it directly at Kimi.com, no setup required.

For developers: call it through the Moonshot AI Open Platform API, which is also what powers Kimi Code when pointed at Kimi models.

Self-hosted: once the full weight release completes, the model can be downloaded and served with standard large-MoE inference frameworks, but 2.8 trillion total parameters demands serious multi-GPU infrastructure, not a personal machine.

Honest Limitations

• Very new: released within the last week, so independent benchmarks and real-world reports are still limited
• Self-hosting is a serious undertaking: this is not a model you run locally without significant hardware
• Vendor benchmarks need verification: Moonshot’s own comparisons are a starting point, not proof
• Weight rollout is still completing: confirm current availability before planning around a full local deployment

Alternatives Worth Knowing

• GLM 5.2, another large open-weight Mixture-of-Experts model with a 1M-token context window
• DeepSeek, a widely used open-leaning model from a Chinese lab
• Poolside Laguna S 2.1, a much smaller open-weight coding model that runs on a single GPU
• Claude and ChatGPT, closed frontier alternatives
• OpenRouter, a way to test Kimi K3 alongside other models through one API

For the broader tradeoffs between models like this and closed platforms, see Open Models vs Closed Models and Frontier Open Models Explained.