An open-weights model with a high MCP Mark that challenges Claude Opus 4.8.
Most of the models in LLMCode Lab share a family resemblance. They are closed, they come from one of the big three labs, and you pay premium per-token rates for the frontier. Kimi K2.7 Code breaks all three assumptions at once, so we gave it a lab of its own at llmcode.ai/kimi.
I had to do my homework on this one. I was excited to explore this model. I will always value the human process.
Here is the short version. It is the only open-weights model in the lab, it scores 81.1 on MCP Mark Verified versus Claude Opus 4.8’s 76.4, it accepts image and video input, and it costs roughly six times less per output token than Opus based on published July 2026 pricing. That combination is unusual enough that it deserved more than a row in a dropdown.
What It Actually Is
Kimi K2.7 Code is Moonshot AI’s coding-focused agentic model, released June 12, 2026 and published under a Modified MIT license on Hugging Face. It is not positioned as a general-purpose chatbot. It is a coding and agent specialist, and it is tuned to behave like one.
Every token wakes up a small slice of a very large model. The router picks which experts handle a given token, so inference stays fast even though the model itself is enormous.
The Number That Matters: Tool Use
Benchmarks are noisy, but one here is worth calling out.
MCP Mark Verified measures how well a model uses tools inside an agentic loop: reading files, running commands, calling APIs, reacting to results.
On MCP Mark Verified, a benchmark for tool use in an agentic loop, Kimi K2.7 Code scores 81.1, while Claude Opus 4.8 scores 76.4.
For an agentic coder, tool use is the job. Reading files, running commands, calling APIs, and reacting to what comes back, that loop is where real coding work happens. On this benchmark, the open-weights and closed-model comparison is worth paying attention to.
The Price Gap Is Not Subtle
Here is how Kimi stacks up against the closest peers already in the lab.
Prices as published by each vendor in July 2026. Anthropic’s first-party pricing page lists Claude Sonnet 4.6 at $3/MTok input and $15/MTok output, so that figure is confirmed. Rates move, so it is worth checking current vendor pricing pages before publishing if this runs later than that.
Compared with Opus 4.8, Kimi is 5.3 times cheaper on input, 6.25 times cheaper on output, and has a higher MCP Mark Verified score, based on published July 2026 pricing and benchmark results. If prompt caching is enabled, Moonshot’s published cached-input price drops to $0.19 per million tokens.
Why a Dedicated Lab Instead of the Main Grid
LLMCode Lab’s main comparison view is built around a four-key, four-provider flow: Hugging Face, OpenAI, Anthropic, Google. Kimi does not fit that mold cleanly, and forcing it in would have complicated the main experience for every user. So we mirrored the pattern already used for the NVIDIA Nemotron lab, a standalone page with its own key entry, model selection, and results. Zero blast radius on the main lab.
A few implementation details that made Kimi its own case:
- OpenAI-compatible endpoint. Moonshot’s API speaks the OpenAI SDK format out of the box at
https://api.moonshot.ai/v1. - Thinking is fixed; sampling guidance is documented. Moonshot’s documentation says K2.7 Code forces
thinkingandpreserve_thinking, and Moonshot’s benchmark notes use temperature set to 1.0 and top_p set to 0.95. So the lab does not expose a temperature slider on the page, and requests follow Moonshot’s documented behavior for the model.
The key lives in your browser and is forwarded to Moonshot only during a request. It does not pass through or get stored on LLMCode Lab’s servers.
As with everything in LLMCode Lab, your API key never touches our servers. It lives in your browser and is forwarded to Moonshot only during a request.
Try It in 30 Seconds
- Grab a key at platform.moonshot.ai.
- Open llmcode.ai/kimi.
- Paste your key, select Kimi K2.7 Code, drop in a prompt, and hit Run.
Give it something agentic, a multi-file refactor description, a gnarly bug with a stack trace, or a “read this schema and write the migration” task. That is the terrain it was built for, and it is the fairest way to see whether the benchmark holds up on your own work.
A live head-to-head, running Kimi side by side with GPT, Claude, and Gemini in one view, is on the roadmap. For now, the standalone lab is the fastest way to put an open-weights frontier coder through its paces.
Worth Keeping an Eye On
An open-weights model posting a higher score than a frontier closed model on an important tool-use benchmark is not a trivial result. It is a sign of where the gap between open and closed is narrowing, and tool use is not a trivial place for that to happen.
If you build with agents, this is worth fifteen minutes of your time in the lab. Subscribe so you catch the head-to-head comparison when it lands, along with everything else running through LLMCode Lab.
Sources: Moonshot AI Kimi K2.7 Code model card and release notes (June 2026); published MCP Mark Verified results; vendor pricing pages checked in July 2026; and LLMCode Lab testing (July 2026).
Reference guide: terms worth knowing
These showed up in the post. Plain language only.
Mixture of experts (MoE)
the architectureInstead of one giant network doing every calculation, the model is split into many smaller specialist networks called experts. A router looks at each token and decides which experts should handle it. Most of the model stays idle at any given moment, which is how a model this large can still run fast.
Active parameters
what actually runsThe portion of the model’s total parameters that actually fire for a given token. Kimi has one trillion parameters total but only about 32 billion active per token, which is why it can be both enormous and fast at the same time.
Context window
how much it can holdHow much text the model can look at in a single pass, measured in tokens, roughly chunks of words. A 256K context window is large enough for large multi-file tasks and substantial portions of a codebase in one pass, so the model can refactor with far more context in view rather than guessing at what it cannot see.
Tool use / MCP
acting, not just answeringThe model’s ability to reach outside itself, reading a file, running a command, calling an API, and then reacting to what comes back. MCP, short for Model Context Protocol, is an emerging standard for connecting a model to those outside tools. This loop is the core of agentic coding work.
Prompt caching
cheaper repeat callsWhen part of a prompt repeats across calls, like a large system instruction or a codebase you keep referencing, the provider can cache it instead of reprocessing it every time. Cached input tokens are billed at a steep discount, which is why Kimi’s effective input cost can drop substantially once caching kicks in.
No terms match that search.
Discover more from MsTechDiva
Subscribe to get the latest posts sent to your email.