Moonshot AI unveils Kimi K3, the world’s largest open-weight AI model: What to know

5 min readNew DelhiUpdated: Jul 17, 2026 02:32 PM IST

Kimi K3 is a 2.8 trillion-parameter model that the Chinese AI startup claims is the world’s largest open-weight AI system. (Express Image/Moonshot)

China appears to be closing the gap with the US in the AI race. Moonshot AI, the Chinese lab behind the Kimi family of AI models, has unveiled Kimi K3, its latest open-weight model. Since its release, the model has attracted considerable attention for delivering performance that approaches leading closed-source systems from US labs.

Kimi K3 is a 2.8 trillion-parameter model that Moonshot says is the world’s largest open-weight AI system. According to the company, it delivers performance close to Anthropic’s Claude Fable and OpenAI’s GPT-5.6. The model also supports a one-million-token context window and native multimodal capabilities, allowing it to process images alongside text.

Kimi K3 uses a mixture-of-experts (MoE) architecture with 896 expert networks, of which only 16 are activated for any given request. This allows it to route tasks to specialised sub-networks instead of engaging its full parameter count each time. It also introduces two architectural innovations – Kimi delta attention and attention residuals – designed to help information persist across the model’s layers. Moonshot says these changes improve scaling efficiency by around 2.5 times compared with Kimi K2, enabling the model to convert additional compute and training data into improved capability more efficiently.

Kimi 3 performance

When it comes to performance, across a set of six widely used coding benchmarks, Kimi K3 is placed first or second in most categories, trailing only Claude Fable 5 in several instances according to Moonshot’s benchmark presentation. It ranked third on Frontier SWE, second on Kimi-internal coding benchmarks and Terminal-Bench 2.1, and first on both ProgramBench and SWE-Marathon. On agentic benchmarks, essentially the tests that measure a model’s ability to follow through multiple steps autonomously rather than just answer single questions, Kimi K3 again ranked at or near the top across eight different evaluations, trailing only Fable 5 and GPT-5.6.

One of the more notable results came from front-end coding evaluations, where Kimi K3 ranked first overall, a jump from 18th place for its predecessor. It topped six of seven measured domains (branding, reference-based design, data analysis, consumer products, simulations, and content tools), finishing second only in gaming-related tasks. In head-to-head comparisons where outputs were judged directly against competing models, Kimi K3 was preferred about 76 per cent of the time on average, compared to roughly 58 per cent for Fable 5 and GPT-5.6.

(Image: Moonshot AI)

The model also performed well outside of coding. On general text and writing evaluations, it moved from 38th to 9th on a combined leaderboard, ranking in the top ten for creative writing, coding, and instruction-following and first in three professional categories: physical and social science, legal and government work, and medicine and healthcare. On an internal editorial-writing benchmark tracked separately, it was reported as the first open-weight model to top that leaderboard, surpassing Fable 5.

On broader “real-world work” indices or benchmarks meant to approximate economically useful tasks such as finance and coding work, Kimi K3 ranked just below Fable 5, ahead of GPT-5.6, Sonnet 5, Opus 4.8, and a recently released Meta model. On a composite index spanning nine evaluations, including Terminal-Bench, Humanity’s Last Exam, and GPQA Diamond, it placed third, behind GPT-5.6 and Fable 5 but notably close to both.

Story continues below this ad

Moonshot demonstrated examples of the model completing large, multi-step agentic tasks such as building playable browser-based games, including 3D exploration games; a Game Boy Advance emulator; and multiplayer-style arena shooters, generating animated motion-graphics explainers, editing a compiled video from dozens of source clips, and even designing a functional computer chip through a multi-hour autonomous run using open-source chip design tools.

Reactions and open questions

Reactions have been mixed. Some observers, including people associated with competing labs, have suggested that the gap between Chinese and American frontier labs has narrowed considerably. Others have pushed back, arguing that flashy demo-style outputs (games, dashboards, and visual UI work) don’t necessarily reflect performance on harder, less visible tasks like navigating and debugging within an existing large codebase.

A separate point of discussion is that Kimi K3’s release did not include a published score on CyberGym, a benchmark used to assess a model’s cyber-offensive capabilities. Scores on this kind of benchmark have previously triggered export restrictions on other frontier models. And, the absence of a published result has led to speculation about whether the model was evaluated on it and what a result might mean for how open-weight models of this capability level are regulated going forward.

A significant part of the discussion around Kimi K3 revolves around cost. It is priced at $3 per million input tokens and $15 per million output tokens. According to Artificial Analysis’ Cost per Intelligence metric, it averages about $0.94 per equivalent benchmark task—similar to GPT-5.6 and roughly half the estimated cost of Claude Opus 4.8 at around $1.80. Separately, Moonshot claims Kimi K3 cost 60–80 per cent less than competing frontier models on several coding benchmarks while matching or outperforming them in some evaluations.