China’s Kimi K3 Claims Open-Weight AI Size Record

Beijing-based AI startup Moonshot AI released Kimi K3 on Thursday 16 July, unveiling a 2.8-trillion-parameter model it says is the world’s largest open-weight AI model. K3 is now available through the Kimi.com chatbot, the Kimi Work desktop agent and Moonshot’s API, arriving just before the 2026 World Artificial Intelligence Conference in Shanghai.

There is an important qualification to Moonshot’s “open” claim. The company has not yet published a K3 checkpoint in its public repository, a licence, a model card or a full technical report. The model is currently accessed through Moonshot’s services rather than downloaded and inspected by developers. Moonshot says its full weights will ship by 27 July, making K3 API-only for now.

A vast model designed to use a fraction of itself

Kimi K3 uses a sparse Mixture-of-Experts architecture, with around 2.8 trillion total parameters. Rather than calling on every part of the model for each request, it activates 16 of 896 experts per token — roughly 1.8% of the available pool. That is intended to give it a vast overall capacity without the full cost of running all its parameters at once.

The model introduces two architectural updates, Kimi Delta Attention and Attention Residuals, which Moonshot says deliver roughly 2.5 times better overall scaling efficiency than its predecessor, Kimi K2. That earlier model had one trillion parameters. K3 also has a one-million-token context window, native visual understanding and support for text, image and video inputs.

Reasoning is always enabled, according to Moonshot, with a tunable reasoning_effort control. Two versions were surfaced at launch: K3 Max for chat and agent tasks, and K3 Swarm Max for large-scale parallel processing. Moonshot described the release as its most capable coding model yet.

“K3 stands as Moonshot AI's most powerful open-source coding model to date.”

Strong early scores, with caveats

Early benchmark results suggest K3 is competitive with leading proprietary systems, although the numbers need careful reading. Moonshot itself says K3 remains behind Anthropic’s Claude Fable 5 and OpenAI’s GPT 5.6 Sol on overall performance, while claiming it beat other models in its own coding and agentic evaluations.

Artificial Analysis placed K3 third on GDPval-AA v2, which measures real-world tasks across 44 occupations and nine industries. Its score of 1,687 put it behind Claude Fable 5 Max and GPT-5.6 Sol Max, but ahead of Claude Opus 4.8. On Artificial Analysis’s private AA-Briefcase agent benchmark, K3 ranked second with 1,527 points.

• Artificial Analysis gave K3 scores of 57.11 on its Intelligence Index, 76.24 for coding and 50.07 for agentic work.
• Arena ranked K3 first in its Frontend Code evaluation at 1,679 points in blind developer testing.
• That represented a 17-place rise from Kimi K2.6, which ranked 18th in the same frontend coding assessment.

These are still early results from a rollout taking place in real time. Some figures are Moonshot’s reported results, while independent testing and documentation are still emerging. The headline claim that K3 “beats” the very best proprietary models is therefore too broad: its top frontend coding ranking is independently reported, but Moonshot says its overall performance still trails the highest-end Claude and GPT systems.

Premium pricing, not a budget model

Moonshot has priced K3 at $3 per million input tokens, $0.30 per million tokens on a cache hit and $15 per million output tokens. That is a sharp increase from the $0.95/$4 per million input/output token pricing cited for Kimi K2.6, and puts K3 closer to the pricing band of Anthropic’s Claude Sonnet than cheaper earlier Kimi offerings.

No official UK price has been confirmed. IBTimes UK estimated a blended rate of about $12, or around £9, per million tokens, but that is not a separately published sterling tariff. Kimi app subscriptions range from $19 to $199 per month.

BofA Securities analyst Alex Liu called K3 “the most expensive Chinese model to date” and said it “raises the capability ceiling for China AI models, shifting the burden of proof to other independent AI labs.”

What happens next

The real test comes on 27 July, when Moonshot says it will release the weights that underpin its open-weight positioning. Even then, K3 will be far from a practical local model for most organisations: Moonshot recommends supernodes with 64 or more accelerators, and has not disclosed the total number of active parameters. In practice, many developers are likely to encounter it through inference providers rather than their own machines.

The launch also marks a striking comeback attempt for Moonshot. DeepSeek’s low-cost R1 model disrupted China’s AI market in January 2025, with Kimi falling from third to seventh in monthly active users in the country. Moonshot responded by pivoting towards open-source models with Kimi K2 in July 2025 and K2.5 in January 2026. K3 is its biggest bid yet to regain attention — and to turn a promised open-weight release into a serious challenge to both Chinese rivals and the biggest US model labs.