AI Radar: Kimi K2.7 Code — One Trillion Parameters, Open, and the Catch in the Benchmarks

One trillion parameters, open weights, and all of it at a fraction of what Claude Opus costs: Kimi K2.7 Code from Moonshot AI is being touted as the affordable, open alternative to Claude Opus. Since June 12, 2026, the weights have been available on Hugging Face. We’ve taken a look at whether the model holds up in everyday use — and where the real catch lies.

Every published figure comes from Moonshot's own benchmark; independent results on SWE-bench Verified or Terminal-Bench were not available at launch. (Quelle: DevOps.com on the K2.7-Code release)

The Key Specs

1 trillion parameters as a Mixture-of-Experts, with 32 billion active per token (384 experts, 8 + 1 shared). The sheer size mainly impacts training; in operation, the model computes with the active 32 billion.
256,000-token context — enough for large codebases and long "plan-and-execute" runs, though not a million-token context like GLM 5.2 or MiniMax M3.
30% fewer reasoning tokens than predecessor K2.6 at higher in-house coding scores — this noticeably reduces cost and latency.
Always thinking: there is no non-reasoning mode; the model "thinks" by default.
Natively multimodal via a MoonViT encoder (text, image, video).
Open weights under Modified-MIT — commercial use and self-hosting are permitted.

The Two Tools Nobody Talks About Much

What’s interesting is less the model alone and more the ecosystem around it. Moonshot delivers two things with it:

Kimi Work — an app interface for knowledge work that lifts the model out of pure chat.
Kimi Code CLI — a terminal-native agent tool in the mold of Claude Code, now with over 6,400 GitHub stars. This is exactly where K2.7 becomes interesting for developer teams: agentic coding directly in the shell, with a model you can self-host if needed.

The Catch in the Benchmarks

Now for the honest part. Moonshot advertises significant jumps — +21.8% on Kimi Code Bench v2, +11% on Program Bench, +31.5% on MLS Bench Lite. The problem: these are all in-house benchmarks. A vendor that designs and evaluates its own tests inevitably optimizes precisely for them — in the scene, this is called “benchmaxing.”

What’s missing is an independent countercheck. K2.7 has not been submitted to established, third-party suites like DeepSWE; on Terminal-Bench, values vary depending on the test harness. Practitioners report a real gap to the top: on a demanding workflow specification, Claude Opus scored 91 out of 100 points, the K2 predecessor only 68 — and the gap lay exactly in the tricky multi-agent and timing errors that rarely show up in benchmark suites.

For fairness, there’s a counterweight: K2.6 once topped the OpenRouter leaderboard — which measures real routing decisions by paying developers, not self-reports. If you use it daily, you clearly see value in it.

Live Test: Two Typical Tasks

In the practical test, the pattern is clear. On finished, well-defined tasks — a small physics simulation or a React pricing component — K2.7 delivers quickly and cleanly. These are exactly the tasks where open models shine in demos: a clear goal, manageable context, immediately visible result.

The reality check comes afterward — in long-lived, nested projects with many dependencies, where an agent must stay consistent for hours. Here the wheat is separated from the chaff, and here the headline benchmarks are least informative. Our recommendation: measure on your own representative repo, not on the demo impression.

Privacy, Censorship, Self-Hosting

Three points need clarification before any productive use:

GDPR: The hosted Kimi API is operated from China. Personal or confidential data thus leaves the EU — for many use cases a deal-breaker.
Censorship: On politically sensitive topics, the model responds with noticeable filtering. For code, this is mostly irrelevant; for general knowledge work, it’s a factor.
Self-hosting: The open weights are the clean escape from the GDPR problem — but they come at a substance cost. Roughly 340 GB and serious GPU resources are needed before the first token is generated.

The Price

The strongest argument remains the price. Via the API, K2.7 costs roughly $0.95 per million input and $4.00 per million output tokens (sometimes cheaper via OpenRouter). Compared to Claude Opus ($5/$25), this is five- to seven-times cheaper — for a model you’re allowed to self-host if necessary.

Our Take

We evaluate models by task, not by brand. Kimi K2.7 Code is a strong, affordable all-rounder for agentic coding — with a serious CLI ecosystem and the freedom of open weights. How its newer successor Kimi K3 stacks up directly against Claude Code and Codex, we examine in our dedicated head-to-head. The benchmark promises should be enjoyed with caution: they come from the vendor itself. If you’re considering K2.7, measure it against your own tasks and clarify the data protection upfront — via self-hosting or an EU-compliant provider. Then it’s one of the most interesting open Claude alternatives the market currently offers.