Kimi K3: China's Open-Weight Model Reaches the Top Tier

What Moonshot AI released

Moonshot AI of Beijing presented its new flagship model Kimi K3 on July 16, 2026, a few days ahead of the World AI Conference in Shanghai. At 2.8 trillion parameters, it is the largest model whose weights are set to become openly available. You can use it today via Kimi.com, Kimi Work, Kimi Code and the Kimi API.

The key facts: a mixture-of-experts design in which only 16 of 896 experts compute per token, a context window of 1 million tokens, native vision, and an always-on reasoning mode. The full weights are scheduled to follow by July 27, 2026 under a Modified MIT license that permits commercial use.

Architecture and benchmarks: what sits behind the numbers

K3 combines two architectural changes: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals. Training was quantization-aware in MXFP4, which makes the weights more compact and later operation cheaper. More important than Moonshot's own figures, though, is what independent testers measure.

The independent tester Simon Willison credits K3 with good vision capabilities and clean code, but points out one quirk: the reasoning mode only knows the max setting. His single SVG test consumed 16,658 output tokens, 13,241 of them for reasoning, and cost 25 cents. How reliable benchmark claims from Chinese labs are in general is something we examined in our piece on China's AI models under benchmark scrutiny. For K3, the central rankings this time come from independent parties, but the weights themselves cannot be verified before July 27.

The end of cheap AI from China

With K3, Moonshot says goodbye to bargain pricing. The API charges 3 US dollars per million input tokens and 15 US dollars per million output tokens, with cache hits at 0.30 US dollars. That is Claude Sonnet's price level. The predecessor K2.6 cost 0.95 and 4 US dollars, so K3 is roughly three to four times more expensive.

Anyone who chose Chinese models mainly for cost reasons needs to recalculate. The always-on reasoning mode sharpens this: even simple requests produce long chains of thought and thus expensive output tokens. A caching strategy and a cost model per use case belong in the design from day one, as described in our piece on token FinOps.

Open weights and self-hosting

The real lever for companies is the weights release on July 27. A frontier-level model you are allowed to run on your own infrastructure changes the negotiating position toward the US providers, even if you never deploy it. The Modified MIT license permits commercial use, adaptation and fine-tuning.

In practice, operating it is anything but trivial. 2.8 trillion parameters mean weights in the terabyte range even at 4-bit quantization. A single server is not enough; a multi-GPU cluster with fast interconnect is realistic. For most teams, the path therefore runs through three stages:

• The Kimi API for first tests with non-critical data, available immediately.
• European GPU cloud providers or managed inference services once the weights are out. For mid-sized companies this is the realistic production route.
• Your own hardware only under permanently high utilization and strict data requirements. What running smaller open models in-house teaches is covered in our piece on local AI models on your own hardware.

European perspective: data protection, AI Act, sovereignty

For European companies, the legal framework decides, not the benchmark. The Kimi API processes data in China, a third country without an adequacy decision. Transfers of personal data then require additional safeguards under Art. 44 ff. GDPR, which is hard to implement cleanly for customer data or confidential documents. The open weights turn the picture around: hosted in-house or in the EU, no data leaves your area of responsibility.

The EU AI Act plays into this as well. K3 is a GPAI model, and the open-source exemption does not apply to models with systemic risk. Whoever runs the weights and substantially modifies them can also take on provider obligations of their own. The deadlines involved are laid out in our overview of the EU AI Act high-risk deadlines.

For Europe's sovereignty debate, the launch stings. Open models at frontier level now come from China, not from Europe. Aleph Alpha and Mistral trail clearly in the rankings, as our piece on Aleph Alpha and Cohere lays out. Anyone who wants to build digital sovereignty on self-hosted models will foreseeably reach for Chinese weights, a dilemma that reshuffles the debate about trustworthy AI.

Challenges and risks

The launch is impressive, but not without open questions. Four points you should know before shortlisting K3.

• Verification is pending. The central rankings are independent, but many detail figures are self-reported. The weights cannot be checked before July 27, and if the release slips, K3 remains an API product for now.
• Only one reasoning setting. The max mode produces long chains of thought even for simple tasks. That drives cost and latency until Moonshot ships graduated modes.
• Governance questions. Training data provenance, built-in content filters, and possible export and sanctions risks of Chinese providers belong in every risk assessment. Our piece on the provenance of AI training data offers a starting point.
• The cost argument no longer carries on its own. At Sonnet-level pricing, K3 competes on capability and control, not on the discount.

What companies should do now

Kimi K3 is a reason to review your model strategy, not a reason for a rushed switch. A staged test under controlled conditions makes sense.

Five priority steps

1. Test with your own benchmarks

   Start the evaluation with non-critical data and measure your own tasks, not the leaderboards. A model that shines on GDPval can still fail on your domain documents.

2. Settle the data protection question first

   Use the API only for workloads without personal or confidential data. For everything else, wait for the weights release and prepare EU hosting options.

3. Work through the cost model

   Reasoning tokens drive the bill. Cache hits cost a tenth of the input price, so a caching strategy belongs in the design, not in the cleanup.

4. Secure your exit capability

   Keep prompts, evals and tooling vendor-neutral. The model market turns every few months; a switch should cost weeks, not quarters.

5. Check your AI Act roles

   Clarify whether self-hosting with fine-tuning makes you a provider under the EU AI Act, and set up your compliance documentation accordingly.

Further reading

Frequently asked questions