Kimi K2.7: The Open-Source Coder That Beat Claude Opus — and the Risk Nobody's Pricing

A trillion parameters, open on Hugging Face

On June 12, Moonshot AI released Kimi K2.7-Code: a one-trillion-parameter coding model with fully open weights, posted to Hugging Face under a modified MIT license. It is a mixture-of-experts design, 32 billion parameters active at any moment, with a 256K context window. On one tool-use benchmark, MCPMark, Moonshot reports 81.1% against 76.4% for Claude Opus 4.8.

Read that last sentence again, because the load-bearing word is “reports.” The number is vendor-supplied, with no independent re-run as of publish. On Moonshot’s own Kimi Code Bench v2 the order flips — 62.0 for Kimi, 67.4 for Opus. “Beats Opus” is a headline that survives exactly one benchmark and one source.

Using a model is not owning the weights

The benchmark is not the story. The license is. There is a real difference between calling a model through someone else’s API and holding the weights yourself, and for a bank, a hospital, or any organization under EU data-residency rules, that difference is the whole game. Open weights mean the model can run inside your own perimeter, on your own hardware, with your code never leaving the building. Reported token costs run roughly an order of magnitude below the closed frontier. That is a genuine promise, and it is the first time a model this capable has come with it.

The newest unmonitored supply-chain input

Here is the part nobody is pricing yet. An open coding agent that writes code which ships is a supplier, whether or not anyone wrote it down. After 25 years in cybersecurity I have learned that the inputs nobody owns are the ones that hurt you. You can audit a vendor. You can question a contract. A set of weights someone pulled from Hugging Face on a Friday answers to no one, and it is now writing a measurable share of your codebase.

Three questions before the weights touch production

One. Who in this organization owns the risk when a model we did not build writes code that reaches production? Two. What did we actually validate about these weights, beyond a benchmark the vendor ran on itself? Three. Where are open-weight agents already running inside our perimeter that no change log ever recorded? The promise here is real. So is the bill that arrives when the answer to all three is silence.