Kimi K3 Now Available in Devin
WindsurfWindsurf's Devin added support for Kimi K3, Moonshot AI's 2.8-trillion-parameter open-weight model, across Devin Desktop, Devin CLI, and Devin Cloud. Cognition calls it the first open-source model it has tested that approaches frontier-level performance, scoring 59.6% on its internal FrontierCode 1.1 benchmark. The model trails Devin's proprietary options (Claude Opus, Claude Fable, and GPT-5.6 Sol) but surpasses GPT-5.5, giving developers a lower-cost, open-weight alternative directly inside the IDE.
Key Takeaways
- Kimi K3 became the first open-source model in Devin to approach frontier-level performance, according to Cognition's own FrontierCode 1.1 benchmark (59.6%).
- The model is available across all three Devin surfaces at once: Devin Desktop, Devin CLI, and Devin Cloud.
- Kimi K3 surpasses GPT-5.5 on Cognition's benchmark but still trails Claude Opus, Claude Fable, and GPT-5.6 Sol, Devin's top proprietary options.
- Cognition flags debugging via code execution as a particular strength, while spec adherence is a relative weakness compared to proprietary models.
- The model uses a sparse mixture-of-experts design (16 of 896 experts active) plus a 1-million-token context window, suited to long agentic coding sessions.
- The release extends Windsurf/Devin's pattern of rapidly expanding its model picker, following recent additions like GPT-5.6 and Claude Opus 5, giving developers more cost/openness trade-offs without leaving the IDE.
Sources & Mentions
3 external resources covering this update
A New Open-Source Option Inside Devin
Windsurf's Devin agent gained a new model option on July 27, 2026: Kimi K3, an open-weight, 2.8-trillion-parameter model from Chinese AI lab Moonshot AI. The model is now live across all three of Devin's surfaces, Devin Desktop, Devin CLI, and Devin Cloud, giving developers a way to run agentic coding sessions on an open-weight model without leaving the Devin ecosystem.
Kimi K3 is built on a mixture-of-experts architecture that activates just 16 of its 896 total experts per task, and it ships with a 1-million-token context window and native multimodal support. Those specs let it hold large repositories and long agent sessions in context, a property Cognition highlighted as important for the kind of long-horizon, multi-step engineering work Devin is designed to automate.
Where It Lands on Cognition's Benchmark
On Cognition's own FrontierCode 1.1 benchmark, which grades real-world engineering tasks on mergeability and code quality rather than pass/fail correctness alone, Kimi K3 scored 59.6%. That places it ahead of OpenAI's GPT-5.5 but behind Devin's three top-tier proprietary options: Claude Opus, Claude Fable, and GPT-5.6 Sol. Cognition specifically framed the milestone as Kimi K3 being "the only open source model that achieves this level of performance" on its benchmark suite.
Strengths and Trade-offs
According to Cognition, Kimi K3 is particularly strong at debugging through actual code execution rather than relying on pattern-matching against training data, a trait that plays well with Devin's sandboxed, tool-using agent loop. Its main weakness, per Cognition's own assessment, is a tendency to deviate from stated specifications more often than the proprietary frontier models, meaning it can wander from precise instructions on more constrained tasks.
Why It Matters
The addition continues a pattern in Devin's recent model roadmap: rather than locking developers into a single frontier model, Windsurf/Devin keeps expanding its model picker with each notable new release, letting developers trade cost and openness against raw capability. Kimi K3's arrival gives cost-conscious teams and developers who prefer open-weight models a credible option inside Devin for the first time, without requiring a separate self-hosting setup.