Kimi K2.7-Code — 81% MCP Score, Zero Independent Benchmarks

Moonshot AI shipped Kimi K2.7-Code on June 12 — a 1-trillion-parameter open-weight coding model that scores 81.1 on something called “MCP Mark Verified,” beating Claude Opus 4.8’s 76.4 on the same suite. It is the first Kimi model to carry the “Code” suffix, and it arrives bundled with a $19/month terminal agent that looks structurally identical to Claude Code. The catch: every single benchmark number supporting those claims comes from Moonshot’s own proprietary evaluation suites. No SWE-bench Verified. No LiveCodeBench. No GPQA Diamond. Nothing an independent researcher can reproduce today.

TL;DR

• What: Moonshot released Kimi K2.7-Code — 1T params, 32B active, 384 experts, 256K context, priced at $0.95/$4.00 per million input/output tokens
• Claim: 81.1% on MCP Mark Verified, beating Claude Opus 4.8’s 76.4% — but the benchmark is Moonshot’s own
• Risk: Zero independent third-party scores exist as of release. Modified MIT license restricts commercial use
• Action: Test as a pricing hedge for agentic pipelines. Do not replace your primary model until SWE-bench Verified numbers land

Kimi K2.7-Code — What Happened

The model itself is technically interesting on paper. K2.7-Code is a Mixture-of-Experts architecture: 1 trillion total parameters, 32 billion active per forward pass, 384 experts, and a 256K-token context window. Moonshot prices it at $0.95 per million input tokens and $4.00 per million output tokens, with cached input at $0.19 per million — roughly 4–5× cheaper per output token than Claude Opus 4.8’s $25 per million.

The performance story centers on four proprietary benchmarks: +21.8% over K2.6 on Kimi Code Bench v2, +11.0% on Program Bench, +31.5% on MLS Bench Lite, and the headline 81.1% on MCP Mark Verified. Moonshot also reports a 30% reduction in reasoning-token usage compared to K2.6, which translates directly to lower cost per task in agentic pipelines that burn through long chains of thought.

Alongside the weights, Moonshot launched Kimi Code CLI — a terminal-first coding agent at $19/month that uses K2.7-Code as its default model. A “6x High-Speed Mode” is listed as coming soon. The model is available on Hugging Face under the ID kimi-k2.7-code and through the Moonshot API.

K2.7-Code ships under a “Modified MIT” license, not standard MIT. The modifications restrict commercial use in ways that standard MIT does not. If you are self-hosting these weights for production agentic workflows, your legal team needs to review the license terms before you treat this as freely deployable infrastructure. Weights on Hugging Face does not mean MIT for all use cases.

Why This Matters

The MCP Mark Verified score is the number Moonshot wants you to focus on, and it is genuinely the most interesting claim in the release. MCP — the Model Context Protocol — is becoming the default execution layer for agentic CI pipelines, and tool invocation accuracy through MCP is one of the hardest things to get right. A model that scores 81.1% on MCP tool use, if that number holds under independent scrutiny, would be the strongest open-weight option for teams building agentic infrastructure on top of MCP servers.

But “if that number holds” is doing enormous work in that sentence. MCP Mark Verified is not the same as MCPMark, the open-source benchmark where K2.6 scored 55.9%. Moonshot’s MCP Mark Verified is a proprietary evaluation suite — different methodology, different test cases, different scoring. When K2.6 launched, its SWE-bench figures were contested. K2.7-Code does not even have SWE-bench figures to contest yet. Moonshot has not submitted K2.7 to any of the industry-standard coding benchmarks as of release day.

This matters because the entire K2.7-Code value proposition rests on claimed performance you cannot independently verify. The pricing is real and verifiable — $4.00 per million output tokens versus Opus 4.8’s $25 is a 6× gap that matters at scale. The 30% reasoning-token reduction over K2.6 is a measurable architectural improvement. But whether K2.7-Code actually produces better code, invokes tools more accurately, or handles complex multi-file edits better than Claude Opus 4.8 — nobody outside Moonshot can tell you that right now.

The broader pattern here is worth naming explicitly. Moonshot is running the exact playbook Anthropic established: ship an open-weight model, build a subscription CLI on top of it, and use benchmarks you control to claim the top spot before independent evaluations exist. Claude Code launched the same way — model plus terminal agent plus curated benchmarks — and it took weeks before the community had enough independent data to form a real opinion. Moonshot is doing this at a fraction of the price point, which makes the lock-in risk lower in dollar terms but higher in switching cost if you build agentic workflows around Kimi Code CLI’s specific patterns.

The 30% reasoning-token reduction is the most concretely valuable claim in this release. For long-horizon agentic tasks — multi-step code generation, iterative debugging loops, autonomous research pipelines — reasoning tokens are the dominant cost driver. Even if K2.7-Code’s absolute quality matches K2.6 rather than exceeding it, the token efficiency alone makes it worth testing in your agentic pipeline.

The $19/month Kimi Code CLI is priced to undercut Claude Code’s $20 Max plan, but the feature gap is unclear. Moonshot lists a “6x High-Speed Mode” as coming soon, which suggests the current speed is not competitive. The terminal agent space is getting crowded — Claude Code, Gemini CLI, Codex CLI — and a dollar cheaper per month is not a moat. The real question is whether Kimi Code CLI’s MCP integration is meaningfully better than the alternatives, and that question loops right back to the unverified benchmarks.

The Take

I would not swap out Claude Opus 4.8 or GPT-5.5 for K2.7-Code in any production pipeline today. The pricing is compelling, the architecture is solid, and the 30% reasoning-token reduction is the kind of concrete improvement that actually moves cost curves for agentic workloads. But every quality claim rests on benchmarks Moonshot controls, and we have been here before with K2.6’s contested SWE-bench numbers.

Here is what I would actually do: run K2.7-Code in parallel on lower-stakes agentic tasks — test generation, documentation updates, refactoring passes — and compare output quality against your current model. If the outputs hold up, the cost savings are real and immediate. If SWE-bench Verified scores land in the next few weeks and confirm the MCP Mark Verified claims, then you have a serious contender for primary model status.

Treat K2.7-Code as a pricing hedge worth testing, not a production replacement. And read the Modified MIT license before you deploy the weights anywhere that matters.

Related

• Kimi K2.6 vs Claude Code — The Cost Threat — How K2.6’s pricing first pressured the Claude Code ecosystem
• Claude Opus 4.8 Dynamic Workflows — The model K2.7-Code is benchmarked against, and what it actually delivers
• The AI Coding Token Cost Reckoning — Why token economics, not benchmark scores, will determine which coding models survive