Moonshot: Kimi K2.7 Code
Coding-specialized MoE model with 1T total parameters, 256K context window, and mandatory thinking mode for long-horizon software engineering. Priced at $0.95 per million input tokens, it executes multi-step agentic tasks with tool-call sequences of 4000+ steps. Built on K2.6 for code generation and debugging across Python, Rust, and Go.
Specifications
| Attribute | Value |
|---|---|
| Lab | Moonshot |
| Tags | Coding Intelligent Agentic Long Context |
| Release Date | 2026-06 |
| Context Window | 262,144 tokens |
| Input Price / 1M | $0.95 |
| Output Price / 1M | $4.00 |
| Input Modalities | Text, Image, Video |
| Output Modalities | Text |
Strengths
- 256K context window for end-to-end code tasks
- $0.95/M tokens input 93% cheaper than GPT-5.5 Pro
- 1T MoE with 32B activated per token
- 4000+ tool-call sequences over 12+ hours
Weaknesses
- Mandatory thinking mode cannot be disabled
- All benchmarks are vendor-reported, no independent scores
- Trails GPT-5.5 and Claude Opus 4.8 on agentic tasks
- Very large self-hosting footprint (~595 GB)
Best For
- Long-horizon coding with deep reasoning chains
- Multi-step agentic workflows using MCP tools
- Code generation across Rust, Go, Python
- Vision-based programming via MoonViT encoder
In Depth: Kimi K2.7 Code
Benchmark Performance
Kimi K2.7 Code coding agent While trailing Claude Fable 5 at 9.9/10, it offers a focused coding specialization at one-tenth the output cost $4.00 versus $50.00 per million tokens.
Moonshot reports vendor-only benchmarks: Kimi Code Bench v2 scores 62.0 (up from 50.9 on K2.6) and MCP Atlas hits 76.0. Independent testers confirm improved code honesty but weaker KernelBench-Hard performance. No SWE-bench scores are available at launch; all figures await third-party verification.
Pricing & Value
Kimi K2.7 Code costs $0.95 per million input tokens and $4.00 per million output tokens. Cache hits drop input to $0.19 making long-context coding affordable for teams.
At $4.00/M output, K2.7 Code delivers 8.5/10 LMRank performance versus Claude Sonnet 5 ($10.00/M output, 9.4/10) for 60% less per token. Compared to GPT-5.5 Pro at $180/M output, K2.7 Code saves 97.8% on generation costs a price-per-point ratio of $0.47 vs $18.95.
Who Should Use This
Professional software engineers building multi-day coding agents. AI teams prototyping long-horizon automation. Researchers needing open-weight access for custom tool chains.
- Agentic developers: build 4000+ step coding workflows for $8.00 per million output tokens.
- Budget-conscious teams: pay 93% less than GPT-5.5 Pro for code-specific MoE inference.
- Open-weight adopters: self-host via vLLM with 1T parameters and 256K context.
- General users: skip this model use K2.6 for writing, analysis, and conversation.
Release & Version History
Released June 12, 2026, Kimi K2.7 Code arrives as a coding-specialized refresh of the K2 series, sitting at LMRank #36 with an 8.5/10 score.
Moonshot AI built K2.7 Code on the K2.6 foundation, activating 32B of 1T parameters per token. The MoE architecture uses 384 experts with 8 selected per token plus 1 shared, MLA attention, and a MoonViT vision encoder. No non-code K2.7 variant exists: the "Code" suffix is essential for correct model selection.