I’ve seen a ton of “open” launches that end up being open in name only: a research license, a download with a gate, weights appearing three weeks after the press cycle. So when Moonshot AI launched Kimi K2.7-Code on June 12, my first question wasn’t about benchmarks. It was: can I download it today, and how much does it cost to run? The answer to both is the real story here, and it’s more interesting than any row on a leaderboard.
Let’s start with what’s verifiable. K2.7-Code is a Mixture-of-Experts model: 1 trillion total parameters, 32B active per token, 384 experts (8 selected, 1 shared), a context window of 256K tokens. The full weights went up to Hugging Face on launch day under a Modified MIT license — the permissive kind that enables commercial use with attribution, not a research-only fence. You can self-host it with vLLM, SGLang, or KTransformers. It comes with native INT4 quantization, and the repo weighs around 595 GB on disk. Let’s be clear: that’s a deployment server-class machine, not something you run on your laptop. But it’s yours to run.
That license and downloadable-today stance are the central point, and worth pausing on. For any team operating under data-residency constraints — and that includes a lot of regulated shops across Latin America — the difference between “API-only, hosted in the US” and “weights you can deploy on your own infrastructure” isn’t a nice-to-have. It’s the line between a tool you can adopt and one your compliance team vetoes. K2.7-Code falls on the right side of that line.
Now the economics. Moonshot prices the API at $0.95 per million input tokens and $4.00 per million output tokens (cached input drops to $0.19). That’s roughly one-tenth what frontier closed models charge. And there’s a second-order effect that matters more than the list price: Moonshot reports the model uses around 30% fewer reasoning tokens than its predecessor, K2.6 — which they frame as “less overthinking.” Since reasoning tokens are billed as output on most price cards, and agentic coding runs hundreds or thousands of steps per task — each plan, retry, and verification paying again to think — a 30% cut compounds hard over a long run. Cheaper per token and fewer tokens per task. That’s the combination that actually moves a monthly bill.
On benchmarks I’m going to be direct: these are Moonshot’s own numbers. The company published a six-row table comparing K2.7-Code against K2.6, GPT-4.5, and Claude Opus 4.8 — and naturally, K2.7-Code wins on every row against its own predecessor, with the headline being a jump from 50.9 to 62.0 on Kimi Code Bench v2 (the +21.8% of the headlines). Against the frontier closed models it’s more honest about the gap: it trails GPT-4.5 and Opus 4.8 on most rows, though it beats Opus on MCP Mark Verified (81.1 vs 76.4). Worth noting that testing conditions weren’t symmetrical — K2.7-Code ran on Kimi’s own CLI, competitors on theirs. Take the table as a vendor’s self-report, not an independent leaderboard. The architecture, the license, and the pricing are the facts I’d base a decision on.
There’s also a platform move underneath the model release that I don’t think is incidental. K2.7-Code launches paired with Kimi Code, Moonshot’s terminal-first coding agent, with membership plans listed from $19/month. So Moonshot competes across the entire stack at once — model, CLI, and subscription economics — the same vertical integration bet the closed players made, except with open weights underneath. That’s a sharper strategy than “here’s a model, good luck,” and it’s the part of this launch I’d be watching over the coming months.
If you’re a CTO weighing your AI coding stack, the takeaway isn’t “switch to Kimi.” It’s that the floor just moved. There now exists an agentic open-weight model, self-hosteable and frontier-adjacent at roughly one-tenth the API cost of closed models, with a license your lawyers can live with. Even if you never deploy it, that’s leverage in every vendor conversation you have this year. Provider freedom stopped being theoretical.
Are you evaluating open-weight models for your coding stack, or is the cost of self-hosting still not working against closed APIs? Tell us how you’re measuring it.