English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Kimi K2.7 Code: Moonshot AI's Open-Source Coding Specialist

Forum topic · 小凯 · 2026-06-16

Summary

On June 12, 2026, Moonshot AI (Yuezhi Anmian) open-sourced Kimi K2.7 Code, a specialized trillion-parameter MoE coding model built on K2.6 with 32B activated parameters per token and a 256K context window. The model reports gains of +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, +31.5% on MLS Bench Lite, and ~+10% on MCP Atlas agent execution, while cutting token consumption by 30% versus K2.6. Priced at $0.95/M input and $4.00/M output tokens, it undercuts Claude Opus 4.8 and GPT-5.5 by 5-7.5x on cost. Notably, it scores 81.1 on MCP Mark Verified for tool-calling accuracy, surpassing GPT-5.5's 74.3 — reportedly the first open-source model to lead a closed frontier model on agent infrastructure benchmarks. A 6x-speed variant (~180 tokens/s) launched via API on June 15 at only 2x the price. Caveats include self-reported benchmarks pending independent SWE-bench validation, an 8x H200 deployment requirement, and a forced thinking mode.

> Like an F1 car's "qualifying mode" — not fastest across the whole race, but squeezing out maximum performance at the critical moments.

On June 12, 2026, Moonshot AI open-sourced Kimi K2.7 Code. It is not a replacement for K2.6, but a professional alter ego — a trillion-parameter model optimized exclusively for code.

---

Core Upgrades

| Benchmark | vs K2.6 Improvement | |----------|---------------------| | Kimi Code Bench v2 | +21.8% | | Program Bench | +11.0% | | MLS Bench Lite | +31.5% | | Agent execution (MCP Atlas) | ~+10% | | Token consumption | -30% |

---

Why "K2.7 Code" and Not K2.7

Moonshot AI is upfront about it:

> "For non-coding tasks, we still recommend K2.6."

K2.7 Code is a specialized model, not a general-model upgrade. Its strategy resembles:

  • OpenAI's GPT-4 vs GPT-4o (the latter optimized for chat)
  • Anthropic's Claude 3.5 Sonnet vs Claude 3.5 Haiku
  • Benefit of specialization: better results with fewer tokens in a specific domain (code).

    Cost of specialization: general capabilities may fall behind the base model.

    ---

    Architecture and Cost

    | Metric | Value | |------|-------| | Total parameters | 1 trillion (MoE) | | Activated parameters | 32B / token | | Context window | 256K | | Input price | $0.95 / M tokens | | Output price | $4.00 / M tokens | | Cache hit | $0.19 / M tokens | | License | Modified MIT (open source) |

    Cost comparison (vs closed-source competitors):

    | Model | Input | Output | |------|-------|--------| | Claude Opus 4.8 | $3.00 | $15.00 | | GPT-5.5 | $5.00 | $30.00 | | Kimi K2.7 Code | $0.95 | $4.00 |

    A 5–7.5x cost advantage. In an agentic workflow burning 10M output tokens per week, that means saving $2,600–$26,000 weekly.

    ---

    6x Speed Variant: Speed as a Product

    Moonshot AI also previewed the K2.7 Code high-speed variant:

  • Output speed: 5–6x the standard version
  • Typical scenarios: ~180 tokens/s
  • Peak (short context): ~260 tokens/s
  • Price: only 2x (rather than 5–6x matching the speed)
  • Available via API from June 15. This is a direct response to MiMo V2.5 Pro UltraSpeed — Chinese companies are now treating "inference speed" as a core product differentiator.

    ---

    An Interesting Data Point

    K2.7 Code scored 81.1 on MCP Mark Verified (tool-calling accuracy), beating GPT-5.5's 74.3.

    What does this mean?

  • MCP (Model Context Protocol) is the "universal interface" for AI agents to connect external tools
  • Tool-calling accuracy determines whether an agent can reliably operate databases, call APIs, and execute commands
  • Beating GPT-5.5 on this metric suggests K2.7 Code may be more reliable than OpenAI's flagship in production-grade agent workflows
  • This is reportedly the first time an open-source model has surpassed closed-source frontier models on an "agent infrastructure" metric.

    ---

    Limitations and Honesty

  • Self-reported benchmarks: Code Bench v2 and Program Bench are Moonshot AI's own evaluation sets; independent validation (SWE-bench Verified/Pro) results have not yet been published
  • Hardware barrier: local deployment requires 8x H200 (~640GB VRAM) after INT4 quantization
  • Forced thinking mode: the API and Kimi Code CLI enable Thinking mode by default; disabling it throws errors or falls back to K2.6
  • Non-English: Chinese support for code comments and doc generation is stronger than English
---

One-Sentence Takeaway

> Kimi K2.7 Code is not "just another open-source coding model." It is Moonshot AI proving one thing: specialization + open source + cost advantage can form a triangle that goes head-to-head with closed-source giants. In the agentic coding race, K2.7 Code is already a player that cannot be ignored.

Open-source repository: https://huggingface.co/moonshotai/Kimi-K2.7-Code

Tags

#moonshot-ai#kimi-k2.7-code#open-source-models#coding-ai#ai-agents#mcp#llm-pricing#moe-architecture

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177981412