Moonshot Releases Kimi K3, the Largest Open AI Model Yet
Kimi K3 broke size records and topped several coding benchmarks. Moonshot's own launch notes say it still trails Claude and GPT on the hardest tasks.
On July 16, 2026, Moonshot AI released Kimi K3, an artificial-intelligence model built on 2.8 trillion parameters, the internal settings a model learns during training. It is the largest openly available model ever published. Within hours, social media had already decided it beat Anthropic’s flagship model, Claude Fable 5. Moonshot’s own launch notes said something calmer: the new model is highly competitive, but there is still a noticeable gap with Claude Fable 5 and GPT 5.6 Sol on the hardest tasks.
For anyone deciding whether to build on Kimi K3, the model card itself is the useful signal, benchmark by benchmark, including the parts where Moonshot admits it still loses.
What Moonshot actually released
Kimi K3 is a mixture-of-experts model, meaning it is not one dense block of parameters but a large collection of smaller specialized ones. Only 16 of its 896 experts activate for any single word it processes, which is why a model this size can still run at a usable speed. It reads and generates text, images, and video through the same underlying network rather than bolting a separate tool onto a text model, and it holds up to a million tokens of context, roughly the length of several long novels, in a single conversation.
The full model weights, the trained parameters other developers can download and run on their own hardware, are not public yet. Moonshot has committed to releasing them by July 27, 2026. Until then, the model is usable through Moonshot’s own products: the Kimi.com chat interface, an agent workspace called Kimi Work, a coding tool called Kimi Code, and a paid API. The official API price is $0.30 per million tokens for a repeated prompt, $3 for a new one, and $15 per million tokens generated, a repeated prompt being one the model has already processed once and cached for a faster, cheaper answer the second time. The release landed the day before the World AI Conference opened in Shanghai, a scheduling choice that reads less like coincidence and more like a company timing its moment.
The gap between what people said and what Moonshot said
Within hours of the release, claims spread on social media that Kimi K3 had simply beaten Fable 5 and GPT 5.6 Sol outright. Moonshot’s own launch page told a more specific story: on its own numbers, the model still shows “a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol” on the tasks that matter most.
A model card that includes the tests it loses is rarer than the marketing around most product launches would suggest. Moonshot published a mixed scoreboard rather than a clean sweep, and that mix is the more useful read than either the hype or the reflexive dismissal that tends to follow a release like this one.
Where Kimi K3 leads
On two of Moonshot’s own coding benchmarks, Program Bench and SWE Marathon, Kimi K3 ranks first, ahead of Claude Opus, both of OpenAI’s GPT models tested, and GLM, a rival Chinese open model. On Terminal Bench it sits essentially tied with GPT 5.6 Sol at the top, ahead of Opus and Fable 5.
An independent measurement points the same direction. LMArena, a site where two AI answers are shown side by side without naming which model wrote which and people vote for the better one, put Kimi K3 in first place for coding, ahead of Fable 5, a jump of 17 places from Moonshot’s previous model in a matter of months. One demonstration that circulated shortly after launch gave both models the same demanding, single-prompt coding task, a real-time simulation with several moving parts, and judged Kimi K3’s result the more technically ambitious of the two, consistent with the coding-specific rankings above even if a single test proves nothing on its own.
Where it still loses
On the Deep SWE and Frontier SWE benchmarks, both Fable 5 and GPT 5.6 Sol come out ahead of Kimi K3. On a broader, general-purpose benchmark the story repeats: Fable 5 leads clearly, though Kimi K3 sits close behind a cluster of other models rather than trailing badly.
There is also a specific technical catch. Moonshot trained Kimi K3 while keeping its full chain of internal reasoning intact between steps, and it has flagged that if the software running the model does not feed that entire reasoning history back in, answer quality can become unpredictable. The company recommends its own tooling for that reason. One demonstration illustrated exactly why: given a pure image-reading test, three identical tanker trucks distinguished only by which one was braking, judged from how the liquid inside was sloshing, Kimi K3 landed on the correct truck only after nearly 20,000 tokens of repetitive internal reasoning, the same question asked over and over before it settled the answer by elimination. The final written answer read as clean and confident, with none of that repetition visible in it, which is the risk in one sentence: a tidy answer can sit on top of an unstable process.
The idea worth keeping
The durable point is that the consensus which formed within hours of the launch outran anyone’s ability to have actually tested the model, while Moonshot’s own published numbers, including the losses, turned out to be the more accurate account. A model can be strong at coding and still trail the leading proprietary models on deep reasoning, and both of those things were true of Kimi K3 on day one.
For anyone weighing whether Kimi K3 belongs in a coding workflow, it is a fair candidate to test specifically for that use case, where its own benchmarks and outside rankings agree it is strong, and where its pricing sits well below what proprietary models like Claude Fable 5 charge. Whether that price gap is worth the trade-off it comes with is its own question, one a direct build test against Fable 5 answers more usefully than either the hype or the denial that arrived on launch day. What “open” actually means for a model like this, and why the license matters more than the parameter count, is covered in more detail here.
FAQ
When will Kimi K3’s weights actually be public? Moonshot has committed to releasing the full model weights by July 27, 2026. Until then, the model is only reachable through Moonshot’s own hosted products and API, not something you can download and run yourself.
Is Kimi K3 better than Claude Fable 5? Not as a blanket claim. It leads on several coding-specific benchmarks and on an independent, human-judged coding leaderboard. On deeper reasoning benchmarks, Moonshot’s own launch notes say Fable 5 and GPT 5.6 Sol are still ahead. Which one wins depends on the task.
How much does Kimi K3 cost to use through the API? Moonshot’s official pricing is $0.30 per million tokens for a cached prompt, $3 per million for a new one, and $15 per million tokens generated. That sits well under what the leading proprietary models charge for equivalent usage.
Is Kimi K3 open source? No. It is open-weight, meaning the trained parameters will be downloadable once released, not open source, which would also mean publishing the training recipe and data. Moonshot has not committed to releasing either of those.
Can a small team realistically run Kimi K3 on its own hardware? Technically, once the weights are out. Practically, no for most teams. A 2.8 trillion parameter model needs data-center-class hardware to run at all, so most people will keep using it through Moonshot’s hosted API rather than self-hosting it.
Kimi K3 is a large jump in what a lab outside the usual handful of American companies has managed to ship, and it wins several coding benchmarks outright. It is not, on Moonshot’s own account, a wholesale replacement for the leading Western models yet. The number worth remembering here is not a leaderboard rank: it is the habit of reading the maker’s own caveats before the takes that arrive within the hour.
