Kimi K3 just became the largest open-weight model ever — here is how to actually use it today

Moonshot AI released Kimi K3 on July 16 — a 2.8 trillion-parameter mixture-of-experts model that activates only 16 of 896 experts per token, with a 1-million-token context window and native image and video input. Three days later, on July 19, Moonshot froze new subscriptions to its own consumer product because demand had risen sixfold in 48 hours and was pushing against the limits of its GPU clusters. Existing subscribers keep their access; anyone who wanted to sign up directly through Moonshot cannot right now.

If you want to try K3 today, that pause does not actually block you. Moonshot's own API and the aggregator OpenRouter both still serve the model — moonshotai/kimi-k3 on OpenRouter, at $3 per million input tokens and $15 per million output tokens, with prompt caching cutting that by roughly 60-80% on repeated context. The catch is capacity, not access: OpenRouter's own listing warns the model "may return frequent 429 errors" while Moonshot works through the backlog. Build in retries before you rely on it for anything production-facing this week.

On the benchmarks, be precise about what "beats Claude" actually means here. K3 took the No. 1 spot on Arena.ai's Frontend Code Arena, ahead of Claude Fable 5 and GPT-5.6 Sol on that specific leaderboard. But Artificial Analysis's broader private evaluation puts K3's Elo at 1547 — second place, trailing Fable 5 overall — and Moonshot's own release notes say K3 sits behind Fable 5 and Sol on aggregate performance while beating everything else, including Claude Opus 4.8 and GPT-5.5. Both things are true at once: genuinely frontier-adjacent on coding-specific tasks, not a frontier-topping model in general.

One real cost gotcha, if you go test it yourself: independent testing (Simon Willison's pelican-SVG benchmark) found K3 currently ships with only one reasoning effort level — "max" — and no way to dial it down. A simple prompt burned over 13,000 reasoning tokens and cost roughly 25 cents to answer. If your workload is mostly short, simple requests, that reasoning overhead will show up in your bill faster than the sticker price suggests. Benchmark your own typical prompt before assuming an open model this size is automatically cheap to run.

On self-hosting: wait for the actual license file. Moonshot has promised full open weights by July 27, but as of this week there is no published Hugging Face repository and no confirmed license text. The K2 line shipped under a Modified MIT license, and most coverage is assuming K3 will follow the same pattern — but that is an assumption, not a confirmed term. If you are planning to self-host K3 for a commercial product, do not commit to that plan until the license file is actually public. A "probably permissive" license is not one you can put in front of your own legal team.

The practical read for builders this week: test K3 through OpenRouter or Moonshot's API now if the Frontend Code Arena result is relevant to what you build (it is a genuinely strong signal for front-end-heavy coding agents specifically); expect rate limits while Moonshot catches up on compute; price out your actual prompt pattern rather than trusting the per-token sticker price; and hold off on any self-hosting commitment until the license lands on or after July 27.