Kimi K3 is a 2.8-trillion-parameter AI model that Moonshot AI, a Beijing-based lab, released on July 17, 2026. It matters for one reason: it is the first Chinese open-weight model to land right next to the top US frontier models on capability, rather than a tier or two behind them. Kimi K3 reads both text and images, handles a 1 million token context window, and undercuts comparable US models sharply on price. This post covers what Moonshot actually shipped, how the benchmark claims hold up, the catch buried in the word "open," and why the reaction has been louder than for any prior Chinese release.
What Moonshot AI actually shipped
Kimi K3 is a mixture-of-experts (MoE) model. Its full size is 2.8 trillion parameters, but it activates only a small slice per token (16 of 896 experts, per Moonshot’s technical write-up covered by MarkTechPost). That design is why a model this large can run at a reasonable cost: most of the network stays idle on any given request.
The headline specs:
- 2.8 trillion parameters: the largest open-weight model announced to date, according to Tom’s Hardware.
- 1 million token context window: enough to hold a large codebase or a stack of long documents in a single session.
- Multimodal: it works across text and images, not text alone.
- Aggressive pricing: $15 per million output tokens, against $50 for Anthropic’s Fable 5. Input is cheaper still, at $3 per million on a cache miss and $0.30 on a cache hit.
The model is live now through Moonshot’s own website and API. That is worth keeping in mind as you read the "open" framing below, because at launch the two are not the same thing.
The benchmark picture, with the fine print
Here is where attribution matters. Most of the eye-catching numbers come from Moonshot itself, and vendor benchmarks always deserve a skeptical read.
On Moonshot’s own reported results, Kimi K3 mostly beats Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5, the tier sitting just behind each lab’s current flagship, while trailing the current leaders, Claude Fable 5 and GPT-5.6 Sol, on overall performance. If you want the plain-language version: K3 is not the best model in the world, but it is close, and it is the best you can point to that also promises open weights.
Some third-party signals back part of that up. The AI evaluation service Arena had Kimi K3 finishing ahead of Fable 5 and GPT-5.6 Sol in a front-end coding test, and ahead of Opus 4.8 in its broader text ranking, per Axios’s reporting on the launch. Coding and general-agent tasks are the areas where K3 looks strongest. General reasoning and the hardest frontier evaluations are where the gap to the leaders is still visible.
Treat all of this as provisional. Independent labs cannot fully verify a model they cannot yet download, and a benchmark score is not the same as reliable behavior in your own production workflow. For context on what "frontier" even means here, we’ve written a separate explainer on frontier models.
The catch: open weights, but not yet
Kimi K3 is described everywhere as open-weight, and that framing is doing a lot of work. Moonshot has said the actual weights will not be published until July 27, 2026. Until then, nobody outside the company can download the model, inspect it, fine-tune it, or run it on their own hardware. You can only reach it through Moonshot’s hosted API.
That distinction matters for a business evaluating K3. Open weights, once released, would let you self-host for data-sensitivity reasons or customize the model for a narrow task. A hosted-only Chinese API is a very different proposition on data governance and vendor risk. For now, K3 sits in the second category, with a promise to move to the first. Note too that "open-weight" is not the same as fully open-source: you get the trained weights, not necessarily the training data or the full recipe.
Why this landed harder than DeepSeek
Chinese open models are not new. We covered DeepSeek’s V4 family and GLM-5.2 from Zhipu as they landed, and both are capable, genuinely cheap open-weight releases. On raw API price, they still undercut K3: DeepSeek V4 runs well under a dollar per million output tokens, and GLM-5.2 sits around $4.40, both below K3’s $15.
So the shock is not price, and it is not the existence of a Chinese open model. It is proximity to the frontier. Earlier releases were clearly a tier or more behind the best US systems. Kimi K3 is being measured against Fable 5 and GPT-5.6 and, on some tasks, winning. That is a different conversation.
The reaction reflected it. Axios ran the headline "China just erased America’s AI lead," and reporting described alarm in both Silicon Valley and Washington. Not everyone agrees the panic is warranted; some analysts argue a single strong release does not overturn the broader US position, and that export controls and compute access still matter. The honest read is that K3 narrowed a gap most people assumed was wider, and did it faster than expected, and that on its own is enough to reset how the next year gets discussed.
What it means for a business choosing a model
The practical signal is not "switch to Kimi K3 tomorrow." It is that the pricing power of the top US labs just took a hit. When a model lands near the frontier at roughly a third of a flagship’s output price, and beats last-generation flagships like Opus 4.8 on coding and agent tasks, the premium you pay for the very best model gets harder to justify for routine work. The pressure is already visible: within a day of K3’s launch, Anthropic reversed course and kept Claude Fable 5 permanently in its Max and Team Premium plans rather than moving it to metered pricing for everyone.
The specific thing worth watching is efficiency under constraint. K3 was trained despite US restrictions on advanced chips, and Moonshot’s whole pitch is that architectural efficiency, not raw compute, got them here. If that holds up once the weights are public on July 27, the assumption that frontier capability requires the largest data centers, and the valuations built on that assumption, is the part of the story with the longest tail. For most businesses, the right move today is to keep your options open, test more than one vendor for your actual tasks, and avoid locking a critical workflow to any single model’s pricing.
Frequently Asked Questions
Who makes Kimi K3?
Kimi K3 is made by Moonshot AI, a Beijing-based AI lab. Kimi is the brand name for its models, and K3 is the third major generation in that line.
When can I download the Kimi K3 weights?
Moonshot has said it will publish the weights on July 27, 2026. Before that date the model is only reachable through Moonshot’s hosted website and API, so you cannot self-host or independently inspect it yet.
Is Kimi K3 better than Claude Fable 5 or GPT-5.6?
On overall performance, Moonshot’s own numbers put K3 behind the current leaders, Claude Fable 5 and GPT-5.6 Sol. On some coding and agent tasks, third-party evaluations from Arena showed K3 ahead of both. It is close to the frontier rather than clearly at the top, and the strongest claims come from the vendor, so treat them cautiously.
How much does Kimi K3 cost to use?
Moonshot lists $15 per million output tokens, versus $50 for Anthropic’s Fable 5. Input is $3 per million on a cache miss and $0.30 per million on a cache hit. It undercuts US frontier models but is more expensive than some other Chinese open models like DeepSeek V4 and GLM-5.2.
What does “open-weight” mean here?
Open-weight means the trained model parameters are released for others to download and run, once Moonshot publishes them. It is not the same as fully open-source, which would typically also include the training data and full training recipe. Open weights let you self-host or fine-tune; a hosted-only API does not.
How is Kimi K3 different from DeepSeek or GLM?
DeepSeek’s V4 models and Zhipu’s GLM-5.2 are also capable Chinese open-weight models, and both are cheaper than K3 on API price. The difference is capability relative to the frontier. Earlier Chinese releases sat clearly behind the top US models; Kimi K3 is being measured directly against them and winning some comparisons.
Should my business switch to Kimi K3?
Not on the strength of launch benchmarks alone. It is worth testing against your actual tasks once the weights are public, but weigh data-governance questions around a China-hosted API and the fact that benchmark scores do not guarantee reliable production behavior. The broader lesson is to test multiple vendors and avoid locking critical work to one model’s pricing.
