Kimi K3 is 'open' the way a container ship is portable

Every few months the open-weights community gets a release that is described as the end of the closed labs' advantage, and every few months the practical reality turns out to be more complicated than the announcement thread. Kimi K3 is the biggest version of both halves of that story: the most impressive open release ever made, and the least runnable.

Moonshot recommends a minimum of 64 accelerators to serve K3. Take the crude four-bit arithmetic — 2.8 trillion parameters at half a byte each — and you are looking at roughly 1.4 terabytes for the weights before you account for activations, KV cache or the million-token context the model advertises. This is not a model you run. It is a model you have a datacentre run for you, which is a different proposition from the one the word 'open' usually implies.

The licence deserves the same scrutiny. A widely circulated claim that K3 ships under a Modified MIT licence is simply wrong; the Hugging Face repository carries a custom Kimi K3 License alongside the 96 weight shards and the technical report. If you are planning commercial deployment, that is a legal review, not a footnote.

Having discounted the hype, the results still deserve respect. K3 took first place in Arena's Frontend Code evaluation at 1,679, ahead of Claude Fable 5, in blind developer testing — that is humans choosing the open model's work without knowing whose it was. It placed second on AA-Briefcase, Artificial Analysis' private long-horizon knowledge-work benchmark, at 1,527, above GPT-5.6 Sol Max at 1,495. And it ranked first on four of eight real-world task-automation benchmarks including SpreadsheetBench 2 and BrowseComp.

But Moonshot also published a comparison table in which its own model loses to Claude Fable 5 on FrontierSWE (81.2 against 86.6), HLE-Full (43.5 against 53.3), OSWorld 2.0, Legal Research Bench — and on Kimi Code Bench 2.0, Moonshot's own coding benchmark, 72.9 against 76.9. You have to respect a lab that ships the table showing it losing on home turf. You also have to read it.

This is where the release bites. K3 lists at $3 per million cache-miss input tokens and $15 output. Claude Fable 5.1 is $10 and $50. Fable wins most of the benchmarks — but the buyer's question is never 'which is better', it is 'is it five times better for my workload'. For a large volume of routine agentic work, the answer is increasingly no, and every open release at this quality drags the acceptable price for good-enough capability further down.

The escalation is also structural rather than one-off. The weekend after K3 landed, Alibaba announced a 2.4-trillion-parameter Qwen 3.8 with open weights — a size class it had previously kept API-only. Two of China's strongest labs moved their largest models into the open within days of each other. That is a strategy, not a coincidence, and it is aimed squarely at the margin the American labs charge for.

One unresolved shadow over all of it: Anthropic accused Moonshot in February of training on 3.4 million Claude exchanges via distillation, and K3 scores within a few points of the models in that complaint. Nobody has proved anything, and distillation claims are notoriously hard to establish. But if you are treating these benchmarks as evidence of independent Chinese frontier capability, that allegation belongs in the same paragraph as the scores — not three pages later.

Two practical weaknesses round out the picture, and Moonshot names them itself. K3 is sensitive to thinking-history truncation, so agent harnesses that trim chain-of-thought degrade its quality — an awkward property for exactly the long-horizon work it scores well on. And it launched with a single reasoning effort level, running at maximum thinking by default; Simon Willison found a single SVG generation test cost 25 cents, which is not the economics anybody expects from an open model. Low and high modes were promised in later updates. Until they land, the cheap-per-token pitch comes with an expensive-per-trivial-request asterisk.