On July 27, 2026, the Chinese lab Moonshot AI published the full open weights of Kimi K3 — a 2.8-trillion-parameter model that is, at the moment of release, the largest openly available AI in the world. 'Open weights' means the model itself is downloadable: anyone with the hardware can run it, fine-tune it and use it commercially without paying Moonshot a cent per token. That is a meaningfully different thing from a free chat tier, and it is why this release matters beyond the usual launch noise.
K3 first appeared for API use around July 16 and became fully open on the 27th. It is a sparse mixture-of-experts model — 2.8 trillion parameters in total, but only a fraction active on any given token, which is how a model this large stays affordable to serve. It handles text and images natively and carries a 1-million-token context window, meaning it can hold something like a small library of documents in mind at once. The interesting question is not whether it is impressive on paper — it is — but who actually benefits from a free frontier-scale model, and who is better served by paying.
Quick summary: Moonshot AI published Kimi K3's open weights on July 27, 2026 (API access opened around July 16). It is a 2.8-trillion-parameter sparse mixture-of-experts model with native vision and a 1-million-token context window — the largest open-weight model available at release. Moonshot's hosted API prices it at about $0.30 per million cached input tokens, $3.00 per million fresh input tokens and $15.00 per million output tokens, flat across the full context. On Moonshot's own reported benchmarks it scores 81.2 on FrontierSWE and 88.3 on Terminal-Bench 2.1 — strong coding numbers, though these are Moonshot-reported and use their own benchmark harness (an independent test measured Terminal-Bench closer to 85), not the industry-standard SWE-Bench Verified. It positions K3 against Claude Fable 5 and GPT-5.6 Sol.
What 'Open Weights' Actually Buys You
There is a real difference between 'free to use' and 'free to own.' A free ChatGPT or Gemini tier is a hosted product: the company runs the model, sets the limits and can change the terms tomorrow. Open weights hand you the model file itself. For a business, that means you can run K3 on your own servers, keep your data entirely in-house, avoid per-token bills at scale and fine-tune the model on your own material. For a developer, it means no vendor can deprecate the model out from under you. That independence is the entire appeal — and it is why open releases from Chinese labs like Moonshot, DeepSeek and Alibaba have become such a competitive force against the closed American frontier.
The catch is hardware. A 2.8-trillion-parameter model is enormous. Even as a sparse MoE that activates only part of itself per token, running K3 locally at full precision requires a cluster of high-end accelerators that costs far more than most individuals or small teams can justify. 'Free' here is free of licensing fees, not free of infrastructure. For the vast majority of people, the practical way to use K3 is through a hosted API — Moonshot's own, or one of the third-party providers that spin these open models up — where you pay per token but skip the hardware entirely.
The Numbers, and How to Read Them
Moonshot reports 81.2 on FrontierSWE and 88.3 on Terminal-Bench 2.1, positioning K3 as a serious coding and agentic model — in the same conversation as Claude Fable 5 (which Moonshot lists ahead of it on FrontierSWE at 86.6) and GPT-5.6 Sol. Two cautions matter. First, these are the vendor's own figures on launch, not independent reproductions, which is normal on day one but not the same as verified. Second, and more subtly: FrontierSWE and Terminal-Bench are not the industry-standard SWE-Bench Verified that most coding comparisons quote. A high score on a lab's own preferred benchmark is a real signal, but it is not directly comparable to a SWE-Bench Verified percentage from another model. We keep them labelled separately for exactly this reason.
| Spec | Kimi K3 | What it means for you |
|---|---|---|
| Parameters | 2.8 trillion (sparse MoE) | Largest open model at release; only part active per token |
| Context window | 1,000,000 tokens | Can hold a small library of documents at once |
| Modalities | Text + vision (images) | Reads screenshots, diagrams and documents, not just text |
| Open weights | Yes, published July 27, 2026 | Self-host, fine-tune, commercial use — no per-token license fee |
| Hosted price (fresh input) | ~$3.00 / million tokens | Cheaper than most closed frontier models per token |
| Hosted price (output) | ~$15.00 / million tokens | Mid-pack; flat across the full 1M window |
| Coding claim | 81.2 FrontierSWE (Moonshot's own benchmark) | Strong, but not the same metric as SWE-Bench Verified |
Should You Actually Use It?
For most individuals, the honest answer is: not directly, and that is fine. You will not download 2.8 trillion parameters onto a laptop. What K3's release does for ordinary users is indirect but powerful — it pushes down the price of capable AI across the whole market. When a frontier-scale model is free to host, every closed vendor faces pressure to justify its subscription, and the cheap hosted tiers get better. The people who benefit directly are companies that want data control, developers who want a model they own, and researchers who want to study a frontier system without a corporate gatekeeper.
- You run a business with privacy or compliance needs: open weights let you keep everything on your own infrastructure — a genuine reason to look at K3.
- You're a developer who's been burned by a deprecated model: an open model you host yourself can't be taken away.
- You're an individual who just wants good answers: you don't need to touch K3's weights — but its existence makes the tools you do use cheaper and better.
- You need verified, apples-to-apples coding benchmarks before betting on it: wait for independent SWE-Bench Verified results rather than launch-day vendor numbers.
01Is Kimi K3 really free?
The weights are free to download and use, including commercially — there is no per-token license fee. But running a 2.8-trillion-parameter model needs serious hardware, so 'free' means free of licensing, not free of infrastructure. Most people will use it through a paid hosted API instead.
02Is it better than ChatGPT or Claude?
On Moonshot's own coding benchmarks it is competitive with frontier models, and it trails Claude Fable 5 on FrontierSWE (81.2 vs 86.6). But those are vendor figures on a non-standard benchmark. Until independent SWE-Bench Verified results exist, treat 'better' as unproven and test on your own tasks.
03What is a 1-million-token context window good for?
It lets the model consider a huge amount of text at once — long books, big codebases, or many documents together — without losing track. It is most useful for research, legal and codebase work where context is large.
04Why do Chinese labs keep releasing open models?
Open weights build developer adoption and global mindshare, and pressure closed Western labs on price. Moonshot, DeepSeek and Alibaba have all used open releases competitively, which is a major reason capable AI keeps getting cheaper.
05Can I try K3 without any setup?
Yes — through Moonshot's hosted API or third-party providers that serve open models, where you pay per token and skip the hardware. That is how most people will realistically use it.
The takeaway on K3 is less about one model and more about the direction of the whole market: frontier-scale capability is now something a lab can give away, and that keeps pushing the real-world price of good AI toward zero. The practical response is to stop marrying a single tool and start comparing. LumiChats puts many current models — open and closed — under one login at a pay-per-day price, so you can run the same real problem through several and see which one actually earns your money.
