Moonshot’s Kimi K3 Raises Stakes in China’s Open-Weight AI Race
The Chinese startup’s 2.8-trillion-parameter model combines near-frontier performance with lower pricing, intensifying pressure on both domestic competitors and leading US developers.
Magnific
Opinions expressed by Entrepreneur contributors are their own.
You're reading Entrepreneur Asia Pacific, an international franchise of Entrepreneur Media.
Chinese artificial intelligence startup Moonshot AI has unveiled Kimi K3. This 2.8-trillion-parameter model, it says, approaches the performance of leading US systems while costing considerably less to run, sharpening China’s challenge to the companies at the frontier of generative AI.
The model is available through Moonshot’s Kimi products and application programming interface, or API. The company plans to release its full weights by July 27, meaning developers will then be able to download and customise the underlying system rather than access it only through Moonshot’s services.
Moonshot describes K3 as the first open model in the three-trillion-parameter class. It has a one-million-token context window and native visual capabilities. It is designed for coding assignments, reasoning and knowledge work that can run over extended periods with limited human supervision.
The launch rattled Moonshot’s listed domestic competitors. Shares of Zhipu AI, which operates under the Z.ai brand, fell 21.9% in Hong Kong by midday on Friday (July 17), while MiniMax dropped 13.8%, Reuters reported.
The market reaction reflects more than the size of the model. Chinese AI companies are increasingly competing on a combination of capability, distribution and price, challenging an industry in which the most advanced proprietary systems have typically commanded substantial premiums.
Moonshot charges $3 per million input tokens and $15 per million output tokens for K3. Anthropic charges $10 for input and $50 for output on Claude Fable 5.
That makes K3 cheaper than the premium US models it seeks to challenge, although it is not inexpensive by market standards. Artificial Analysis described K3 as somewhat costly compared with other models in a similar price range and found it to be slower and more verbose than the median system in its peer group.
Moonshot said K3 performed competitively with Claude Fable 5 and surpassed Claude Opus 4.8, GPT-5.6 Sol and GPT-5.5 in an internal test of GPU-kernel optimisation, a process used to improve how efficiently AI software runs on specialised chips. The company said all models were tested in identical computing environments, although a third party supplied the Fable 5 result and may have involved fallback behaviour.
Independent tests have also placed K3 among the leading systems, though early results should be treated cautiously. Arena.ai ranked it first on its preliminary WebDev leaderboard, which measures performance on front-end development tasks. Vals AI placed K3 second on its broader index, behind Fable 5, while Artificial Analysis gave it a score of 57 on its Intelligence Index.
Benchmarks can vary according to the tasks, software tools and computing resources used, and they do not always predict how reliably a model will perform in commercial deployments. Moonshot itself said K3 remained behind Fable 5 and GPT-5.6 Sol overall, even as it outperformed leading systems on some coding and agent-based evaluations.
The headline parameter count also requires qualification. K3 uses a mixture-of-experts architecture, dividing the model into specialised components and activating 16 of its 896 experts for a given task rather than deploying all 2.8 trillion parameters at once. Moonshot said architectural and training improvements made K3 about 2.5 times more efficient at converting computing power into performance than its Kimi K2 predecessor.
The release comes as Chinese developers accelerate the pace and scale of new models. Z.ai’s GLM-5.2 has performed close to leading US proprietary systems on some benchmarks, while MiniMax is developing a 2.7-trillion-parameter model that could be released as early as the third quarter of 2026, Reuters previously reported.
Moonshot, founded in 2023, is backed by investors, including Alibaba and Tencent. The company was seeking about $2 Bn in fresh capital at a valuation of roughly $30 Bn ahead of a possible Hong Kong listing, Reuters reported, citing Bloomberg.
K3 does not establish that China’s AI companies have overtaken the leading US laboratories. Its weights and complete technical report have yet to be published, and Moonshot acknowledges that the model still trails the strongest proprietary systems overall.
What the launch changes is the commercial pressure around frontier AI. As Chinese developers offer increasingly capable models at lower prices and with fewer deployment restrictions, the contest is becoming less about who briefly leads a benchmark and more about which systems developers and companies ultimately choose to build on.
Inputs from Reuters
Chinese artificial intelligence startup Moonshot AI has unveiled Kimi K3. This 2.8-trillion-parameter model, it says, approaches the performance of leading US systems while costing considerably less to run, sharpening China’s challenge to the companies at the frontier of generative AI.
The model is available through Moonshot’s Kimi products and application programming interface, or API. The company plans to release its full weights by July 27, meaning developers will then be able to download and customise the underlying system rather than access it only through Moonshot’s services.
Moonshot describes K3 as the first open model in the three-trillion-parameter class. It has a one-million-token context window and native visual capabilities. It is designed for coding assignments, reasoning and knowledge work that can run over extended periods with limited human supervision.
