By Moonshot AILearning Objectives
- Understand what Kimi Code is and its relationship to the Kimi K2.6 and K2.7 Code models
- Evaluate Kimi Code's benchmarks, pricing, and open-source advantages
- Assess Moonshot AI's explosive growth as China's fastest AI decacorn
What Is Kimi Code?
Kimi Code is an open-source AI coding agent from Moonshot AI, launched January 27, 2026 alongside the Kimi K2.5 model. Its engine has advanced quickly: the CLI moved to Kimi K2.6 in spring 2026 and is now powered by Kimi K2.7 Code, a coding-tuned model Moonshot released on June 12, 2026. Built on the K2.6 base, K2.7 Code is a 1 trillion parameter mixture-of-experts (MoE) model — 32 billion parameters active per token across 384 experts — with a 256,000 token context window, and it ships its open weights on Hugging Face under a Modified MIT license. The agent's tooling, IDE integrations, and CLI surface are unchanged; the underlying model is simply faster and stronger.
Think of it as China's answer to Claude Code: a terminal-first coding agent that can plan, write, debug, and test code autonomously, with IDE integrations for VS Code, Cursor, and Zed. It is fully open-source under Apache 2.0 with 6,400+ GitHub stars — making it the strongest open-source coding agent available.
💡Key Concept
Terminal-First Coding Agent: Like Claude Code, Kimi Code runs primarily in your terminal — you describe what you want built, and it writes code, creates files, runs commands, and debugs errors. The terminal-first approach means it works with any project setup and integrates naturally into existing developer workflows without requiring a specialized IDE.
Key Capabilities
- Terminal CLI agent — plan, write, debug, and test code from natural language descriptions
- IDE integrations — VS Code (extension), Cursor, and Zed (via Agent Client Protocol)
- MCP support — full Model Context Protocol tool integration for connecting to external services
- Dual mode — coding agent mode and shell mode (toggle with Ctrl-X)
- Multimodal input — accepts images and videos (screenshot-to-code workflows)
- 256,000 token context window — handles large codebases and long conversations
- 1 trillion parameter MoE architecture — 32 billion parameters active per token across 384 experts, balancing capability with efficient inference
- Roughly 30 percent fewer reasoning tokens — K2.7 Code reaches its answers with less intermediate "thinking" than K2.6, cutting latency and cost
- 100 tokens per second output — fast generation speed
- Agent Swarm (research preview) — concurrent task execution via Parallel-Agent Reinforcement Learning (PARL), up to 4.5 times faster execution
Benchmark Performance
Moonshot reports that K2.7 Code posts double-digit gains over the previous K2.6 model on its coding-focused evaluations:
| Benchmark | K2.6 → K2.7 Code | What It Measures |
|---|---|---|
| Kimi Code Bench v2 | 50.9 → 62.0 (+21.8%) | Real-world coding-agent tasks |
| Program Bench | 48.3 → 53.6 (+11.0%) | General programming ability |
| MLS Bench Lite | 26.7 → 35.1 (+31.5%) | Machine-learning engineering tasks |
Moonshot benchmarked K2.7 Code against GPT-5.5 and Claude Opus 4.8, and highlights that it edges out Opus 4.8 on the MCP Mark Verified tool-use evaluation (81.1 versus 76.4) while using roughly 30 percent fewer reasoning tokens. For context, the prior K2.5 generation scored 76.8 percent on SWE-bench Verified — open-source state of the art at its launch.
⚠️Warning
Read vendor benchmarks with care. These figures are Moonshot's own reported results. After K2.7 Code shipped, some practitioners told VentureBeat the headline gains were hard to reproduce in everyday use — a recurring pattern as open-source labs race to claim coding leadership. Treat the numbers as a directional signal and test on your own workloads before switching.
Pricing
At under one dollar per million input tokens, K2.7 Code's API stays aggressively priced for its tier — still far cheaper than most Western coding models, even after the increase from the K2.5 and K2.6 generations. Cached input drops to roughly 19 cents per million tokens for repeated context.
Kimi Code vs. Competitors
| Tool | License | SWE-bench | Key Difference |
|---|---|---|---|
| Kimi Code | Modified MIT (open weights) | K2.5: 76.8% | Free CLI; aggressive API pricing; leading open-source coding agent |
| Claude Code | Proprietary | ~80.9% (Opus) | Higher benchmark scores; backed by Anthropic; subscription-based |
| GitHub Copilot CLI | Proprietary | Not published | Largest user base; GitHub integration; $10-19/month |
| Devin | Proprietary | 13.86% | Fully autonomous agent; browser + terminal; $20/month + ACU costs |
| Augment Code | Proprietary | 70.6% | Enterprise focus; Context Engine; SOC 2 + ISO 42001 |
Kimi Code's advantage: The only top-tier coding agent that is fully open-source (Apache 2.0) with competitive benchmarks. You can self-host it for free or use the API at $0.60 per million input tokens.
⚠️Warning
Kimi K2.5 has faced controversy over allegations of not crediting open-source foundations used in its development. If open-source provenance and attribution matter to your organization, review the model's training disclosure before adopting.
Company Details
| Detail | Info |
|---|---|
| Company | Moonshot AI |
| Founded | March 2023 |
| CEO | Yang Zhilin (age 33; NLP researcher with transformer architecture background) |
| Headquarters | Beijing, China |
| Employees | ~80-300 |
| Total Raised | ~$1.77 billion across 3 rounds |
| Latest Funding | $2 billion round (May 2026; led by Meituan's Long-Z Investments) |
| Valuation | $20 billion (May 2026) |
| Revenue | $240 million+ (2025); 20-day post-K2.5 revenue exceeded entire 2025 annual total |
| Growth | Overseas API revenue quadrupled; monthly paying users up 170%+ |
| Status | China's fastest company to reach decacorn ($10 billion+) status |
| Website | kimi.com/code |
Strengths
- Open weights (Modified MIT) — free to self-host via vLLM, SGLang, or KTransformers; no vendor lock-in
- Strong reported benchmarks — double-digit K2.7 Code gains over K2.6 on Moonshot's coding evaluations; edges out Claude Opus 4.8 on its MCP tool-use test
- Aggressive API pricing — under one dollar per million input tokens (about 95 cents on a cache miss, 19 cents cached); still far cheaper than most Western alternatives
- Broad IDE support — VS Code, Cursor, Zed, and terminal CLI with MCP integration
- Explosive company growth — $240 million+ revenue; China's fastest decacorn; backed by Alibaba and Tencent
Limitations and Considerations
- China-based — data privacy and regulatory considerations for organizations in regulated industries
- Attribution controversy — allegations of insufficient crediting of open-source foundations in K2.5 development
- Trails frontier models — 76.8% SWE-bench vs. Claude Opus at 80.9% and Qwen3-Max at 88.3%
- Newer product — launched January 2026; less battle-tested than GitHub Copilot or Cursor
- English documentation — while the CLI works globally, some documentation and community resources are primarily in Chinese
Key Takeaways
- Kimi Code is a leading open-source coding agent — a CLI with VS Code, Cursor, and Zed integrations and MCP support, now powered by the Kimi K2.7 Code model (June 2026) under a Modified MIT open-weights license
- K2.7 Code is a 1 trillion parameter mixture-of-experts model (32 billion active, 384 experts) with a 256,000 token context window, built on Kimi K2.6; Moonshot reports gains of about 22 percent on its Kimi Code Bench v2 and 32 percent on its machine-learning engineering benchmark, with roughly 30 percent fewer reasoning tokens
- API pricing is about 95 cents per million input tokens (19 cents cached) and four dollars per million output — a step up from the prior generation but still aggressive for the tier
- Independent reviewers have cautioned that Moonshot's reported benchmark gains can be hard to reproduce in everyday use, so test on your own workloads before switching
- Best for developers wanting a fast, low-cost, open-weights alternative to Claude Code with strong tool-use performance
