Kimi K3, a 2.8T open model, lands in GitHub Copilot
Every paid Copilot tier can now select Moonshot's model in-IDE — near-frontier coding at a fraction of closed-model API prices; team admins must enable it first.
Copy markdown
Pick it from the Copilot model dropdown today
Kimi K3 is now selectable on Copilot Pro, Pro+, Max, Business, and Enterprise — hosted by GitHub on Fireworks AI and usable in VS Code, Visual Studio, JetBrains, Xcode, Eclipse, the Copilot CLI, and the cloud agent. It's rolling out gradually, and Business/Enterprise admins have to flip it on: it's off by default.
The model: 2.8T params, open weights, 1M context
K3 is a Stable-LatentMoE design — 2.8 trillion total parameters with just 16 of 896 experts firing per token — plus a 1M-token window and native multimodal input. Moonshot published the full weights in late July, so you're not locked to Copilot: self-host it, or hit the API at $3 in / $15 out per million ($0.30 cached, with 90%+ cache hits on code).
Closed-model coding, open-model price
K3 sits top-5 on public coding and agentic leaderboards and trades blows with Claude Fable 5 and GPT-5.6 Sol, while undercutting the closed flagships it competes with on token cost. That price gap is reportedly why Microsoft trialed it internally to cut Copilot's inference bill before shipping it to everyone.