Image: GitHub GitHub Copilot Adds Kimi K3 and Microsoft's Own MAI-Code Model
Two new models landed in GitHub Copilot this week: Kimi K3 from Moonshot AI at $3 per million tokens, and MAI-Code-1.1-Flash, Microsoft's own code model with native image understanding. Agent Plugins 1.0 also went generally available.
GitHub’s weekly release notes for August 10 included two model additions and a platform milestone that together signal a busier week than usual for Copilot.
Kimi K3 is now generally available. Moonshot AI’s open-weight model joined GitHub Copilot on August 6, though the rollout was briefly paused to address a separate GitHub Actions incident before resuming. Kimi K3 shows up in the model picker across all Copilot plans (Pro, Pro+, Max, Business, Enterprise) and works in VS Code, Visual Studio, Copilot CLI, the cloud agent, GitHub.com, GitHub Mobile, JetBrains, Xcode, and Eclipse. Pricing is $3 per million input tokens, $15 per million output tokens, and $0.30 per million cached input tokens.
The model is positioned as a cost-effective option for agentic coding tasks. Moonshot markets it as frontier-level on software engineering benchmarks at a price point below comparable models.
MAI-Code-1.1-Flash is rolling out. This is Microsoft’s own code model, distinct from the OpenAI models Copilot has historically relied on. It ships with native image understanding, which means it can process screenshots, diagrams, and UI mockups alongside code. Microsoft says the model also improves on instruction following, tool use, and coding quality over the previous version. This is a notable step: Microsoft building and shipping its own code-specific model rather than depending entirely on OpenAI.
Agent Plugins 1.0 went generally available. Plugins built once now work across Copilot CLI, the VS Code agent, the cloud agent, and the Copilot SDK. This is the platform play: a developer builds an integration and it shows up everywhere in the Copilot ecosystem without additional adaptation.
Other additions in the August 10 weekly release:
- CLI: a
/taskscommand to manage subagents, the ability to queue prompts during agent execution, and a/rewindcommand to restore file changes without going through git - JetBrains: Copilot memory for maintaining context across sessions, plus Ollama support as a bring-your-own-key provider
- VS Code 1.133: model switching between Claude BYOK and built-in Copilot models, pinned prompts in long chats, and live HTML preview without a manual refresh
The Microsoft-built MAI-Code model is the most interesting development in the batch. Copilot has always been positioned as an OpenAI partnership product. Shipping a Microsoft-authored model changes that positioning, at least at the margins.
Sources: