Kimi K3 is an open-weight model from Moonshot AI. At 2.8 trillion parameters it is the largest model ever released openly, it trades first and second place with Anthropic’s Fable 5 on coding benchmarks, and it runs at roughly a third of the price. It sits fourth in the world on the broadest independent benchmark, under three points behind the best model on Earth, and you can download it.
What is Kimi K3?
Moonshot AI is one of the leading labs in China, backed by Alibaba, Tencent and Meituan. It sits alongside ZAI, which released GLM 5.2 a few weeks earlier. Fable 5, the most capable model in the world today, launched in early June 2026. K3 landed five weeks after it, in July.
Start with size. K3 runs on 2.8 trillion parameters. DeepSeek’s flagship sits around 1.6 trillion and GLM 5.2 runs on 744 billion, so K3 is roughly 75% larger than DeepSeek and close to four times the size of GLM.
Its context window holds over a million tokens, level with Claude and GPT. It is natively multimodal, so text, images and video all go in as input.
Kimi K3 benchmarks: how it scores against Fable 5 and GPT 5.6 Soul
| Benchmark | Kimi K3 | Where Fable 5 and GPT 5.6 Soul sit |
|---|---|---|
| Artificial Analysis Intelligence Index | 4th, scoring 57.1 | Fable 5 and two configurations of Soul are the only models above it, all inside three points |
| Valse AI, 38 models | 2nd | Fable 5 1st |
| SWE marathon (long-running engineering work) | 1st | both behind K3 |
| Program Bench | 1st | both very close behind |
| Terminal | 2nd | Soul 1st, by half a point |
| Frontier SWE | 2nd | Fable 5 1st by over five points, Soul about 15 points behind K3 |
| Arena front-end coding (humans vote) | 1st, winning 76% of match-ups | both behind K3 |
The Intelligence Index blends nine tests into a single score for general capability. Below K3 sit Opus 4.8, GPT 5.5 and GLM 5.2, so an open model you can download is already beating the main production models from both Anthropic and OpenAI. Valse AI, a second independent tracker, put it second out of 38.
Coding is what the model was built for, and the pattern there is that K3 trades first and second place with Fable 5 and Soul depending on which test you pick. On Frontier SWE, a harder software engineering test, Fable 5 beats it by over five points, but K3 still finishes about 15 points ahead of Soul and Opus. It consistently beats Opus 4.8.
The Arena front-end leaderboard is the result to read twice, because real people vote on which model built the better thing rather than a script marking a test. K3 took first place, ahead of Fable 5 and Soul, winning 76% of its match-ups. Its previous version ranked 18th.
How much does Kimi K3 cost?
$3 per million tokens in and $15 per million out. Fable 5 charges $10 in and $50 out, so K3 does the same work for about 70% less. Claude Sonnet 5 is priced level with K3.
| Model | Parameters | Per million in | Per million out | Open weights |
|---|---|---|---|---|
| Kimi K3 | 2.8 trillion | $3 | $15 | Yes |
| Fable 5 | $10 | $50 | No | |
| Claude Sonnet 5 | on par with K3 | on par with K3 | No | |
| GLM 5.2 | 744 billion | Yes | ||
| DeepSeek flagship | around 1.6 trillion | Yes |
Put a real month through it. Say your after-hours intake agent reads 10 million tokens of call notes, job history and price lists, and writes 10 million tokens of replies and bookings back out. On K3 that is $30 in and $150 out, so $180 for the month. On Fable 5 it is $100 in and $500 out, so $600. Same job, $420 a month of difference, and output is the expensive half on both.
How to use Kimi K3
Three ways in, none of which need any setup. Go to kimi.com and chat with it in the browser. Use the API when you are plugging K3 into something you are building. Or use Kimi Code, their coding agent, in the same shape as Claude Code and Codex.
Is Kimi K3 open source?
It is open weights, which is the more precise term. You get the model file itself, free to download, the same arrangement as Llama and GLM 5.2. A closed model like Fable 5 never leaves the lab’s servers. You rent access, they keep the keys, and access can be taken away.
Open weights is a narrower thing than open source in the software sense. Publishing the weights does not mean publishing the training data or the code that produced them, and the licence attached decides what you are allowed to build. Read the licence before you put a downloaded model under something you sell.
Can you self-host Kimi K3?
Free to download and free to run are different problems. K3 is close to four times the size of GLM 5.2 and 75% larger than DeepSeek’s flagship, and all 2.8 trillion parameters have to sit on hardware for the model to answer a single question. That is a data centre, not the machine on your desk.
If you run a solar or HVAC business you are never going to host 2.8 trillion parameters next to the dispatch board. Use kimi.com or the API. Both send your data to Moonshot, a Chinese company, and whether that is acceptable depends on what you are sending. Call notes and customer addresses are a different question from a draft of a marketing email.
What Kimi K3 means for what you pay for AI
Not long ago the estimated gap between the best American model and the best Chinese model was seven months. Fable 5 launched in early June. K3 landed five weeks later and is going toe-to-toe with it as an open model.
A frontier model being open weights changes the economics for everybody. When the best thing available is free to download and a third of the price to run, every other lab has to respond, and the responding is done with price. More competition means more capability for less money.
So whatever you are paying per million tokens today, do not sign a two-year deal on it. Ask whoever sells you an AI system which model is underneath, and what happens to your invoice when a cheaper one that scores the same comes out. That question is the difference between paying for tokens and paying for a subscription.
