Overview
Kimi K2.7-Code is Moonshot AI's coding-focused agentic model, built on top of Kimi K2.6. It is a 1-trillion-parameter Mixture-of-Experts model that activates 32B parameters per token (384 experts, 8 selected per token plus 1 shared expert, across 61 layers), and ships with open weights under a Modified MIT License.
The model targets real-world, long-horizon software engineering: multi-step coding agents, codebase understanding, and tool-driven workflows. Moonshot reports substantial gains on long-horizon coding tasks over K2.6, together with roughly a 30% reduction in thinking-token usage, which lowers the cost of agentic runs.
K2.7-Code is served through Moonshot's Kimi API platform with OpenAI- and Anthropic-compatible endpoints, supports a 256K-token context window and tool calling, and runs with thinking enabled. It is also the model behind the Kimi Code CLI for terminal and IDE coding agents.
| Released | 2026 |
|---|---|
| License | Modified MIT License |
| Weights | Open weights |
| Parameters | 1T total · 32B active |
| Context | 256K |
| Architecture | Mixture-of-Experts |
| Modalities | Text, Vision, Video |
| Status | Available |
Benchmarks
Evaluation Results — Kimi K2.7 Code vs. Kimi K2.6, GPT-5.5, and Claude Opus 4.8 on coding and agentic benchmarks (thinking mode enabled via Kimi Code CLI, temperature = 1.0, top_p = 0.95, averaged across multiple runs).
| Benchmark | Kimi K2.6 | Kimi K2.7 Code | GPT-5.5 | Claude Opus 4.8 |
|---|---|---|---|---|
| Kimi Code Bench v2 | 50.9 score | 62 score | 69 score | 67.4 score |
| Program Bench | 48.3 score | 53.6 score | 69.1 score | 63.8 score |
| MLS Bench Lite | 26.7 score | 35.1 score | 35.5 score | 42.8 score |
| Kimi Claw 24/7 Bench | 42.9 score | 46.9 score | 52.8 score | 50.4 score |
| MCP Atlas | 69.4 score | 76 score | 79.4 score | 81.3 score |
| MCP Mark Verified | 72.8 score | 81.1 score | 92.9 score | 76.4 score |
This model's scores
- Kimi-Code-Bench v262%
- Program-Bench53.6%
- MCP Mark Verified81.1%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Pricing
| Input | $0.95 / 1M tokens |
|---|---|
| Cached input | $0.19 / 1M tokens |
| Output | $4.00 / 1M tokens |
Strengths
- Strong long-horizon, multi-step agentic coding performance built on the K2.6 base.
- Token-efficient reasoning — about 30% fewer thinking tokens than K2.6, cutting agent run costs.
- Large 256K-token context for whole-codebase understanding and long trajectories.
- Native tool calling with OpenAI/Anthropic-compatible APIs for easy agent integration.
- Open weights (Modified MIT) enabling self-hosting and customization.
Best for
- Autonomous coding agents that plan and execute multi-step software tasks.
- Code generation and refactoring across large, multi-file repositories.
- Terminal/IDE coding assistants via the Kimi Code CLI.
- Tool-using agents that integrate external systems through MCP-style workflows.
How to access
| Provider | Model ID |
|---|---|
| Moonshot AI (Kimi) ↗ | kimi-k2.7-code |
Kimi K2 — every version
The full lineage of the Kimi K2 line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Kimi K2.7-Codecurrent | 2026-06-12 | 256K | Modified MIT |
| Kimi K2.6 | 2026-04-20 | — | Open weights |
| Kimi K2.5 | 2026-01-27 | — | Open weights |
| Kimi K2-Instruct-0905 | 2025-09-09 | — | Open weights |
| Kimi K2 | 2025-07-11 | — | MIT |
FAQ
What is Kimi K2.7-Code?
Kimi K2.7-Code is Moonshot AI's open-weight, coding-focused agentic model built on Kimi K2.6. It is a 1-trillion-parameter Mixture-of-Experts model that activates 32B parameters per token, tuned for long-horizon software engineering, code generation, and tool-using coding agents, and released under a Modified MIT License.
How large is the context window?
Kimi K2.7-Code supports a 256K-token context window, large enough to load substantial portions of a codebase and to sustain long, multi-step agentic coding trajectories. Combined with native tool calling, this makes it well suited to whole-repository understanding and extended autonomous engineering sessions.
How much does the Kimi K2.7-Code API cost?
On Moonshot's Kimi platform, K2.7-Code is priced at $0.95 per million input tokens and $4.00 per million output tokens, with cache-hit input billed at $0.19 per million tokens. The model also reduces thinking-token usage by roughly 30% versus K2.6, which lowers the effective cost of agentic runs.
Are the model weights open?
Yes. Kimi K2.7-Code is an open-weight model published on Hugging Face under a Modified MIT License, so it can be self-hosted and customized. It is also available as a hosted API on Moonshot's Kimi platform with OpenAI- and Anthropic-compatible endpoints for teams that prefer managed inference.