Moonshot launched Kimi K3, a 2.8 trillion-parameter open-weight AI model approaching the performance of Anthropic's frontier Fable system
In Hong Kong, the announcement caused shares of domestic competitors Zhipu and Minimax to drop by 27.7% and 16.5%, respectively
Kimi K3 features a one million-token context window designed for advanced reasoning, long-horizon coding, and knowledge work
Chinese AI startup Moonshot unveiled Kimi K3 on Friday, launching a 2.8 trillion-parameter model, according to a Reuters report. Moonshot stated the system is the world's largest open-weight AI model and approaches the performance of US developer Anthropic's frontier Fable model.
The launch arrives a month after the US government abruptly withdrew Anthropic's Fable and Mythos models due to security concerns.
4 July 2026
Get the latest issue of Outlook Business
The announcement immediately impacted the markets. Shares of domestic AI competitors Zhipu (2513.HK) and Minimax (0100.HK) dropped 27.7% and 16.5%, respectively, in Hong Kong just before market close.
Technical Capabilities and Benchmarks
Moonshot said Kimi K3 is the first open-weight model to approach the 3 trillion-parameter mark. It features a 1 million-token context window designed for advanced reasoning, long-horizon coding and knowledge work.
Open-weight models allow users to download, run and customise the systems, unlike closed-source proprietary models.
Moonshot stated Kimi K3 "performed competitively with Fable 5 (with fallback) and substantially outperformed Anthropic's Opus 4.8, GPT 5.6 Sol, and GPT 5.5" in GPU kernel optimisation, which refers to techniques that maximise AI hardware utilisation and minimise latency.
The company said the model incorporates two significant architectural upgrades that improve computing efficiency and enable it to complete long-horizon coding tasks with minimal human supervision.
Third-party evaluations confirm strong performance. Arena.ai ranked Kimi K3 first in web interface-building. Vals AI placed the model second overall behind Fable 5 and ahead of GPT-5.6 Sol. Artificial Analysis reported the model is comparable to GPT-5.5 and Claude Opus 4.8 on complex multi-step tasks, as per Reuters.
Cost and Deployment Realities
Companies including Moonshot, Z.ai and MiniMax are releasing powerful models at sharply lower costs. This trend directly challenges Western assumptions that Chinese developers trail their American peers by months.
Omdia chief analyst Lian Jye Su said Chinese models "can be run at a fraction of the cost that US AI charges its clients". However, he added that massive scale "doesn't necessarily mean you have the best performance by default."
These economic advantages carry heavy hardware requirements. Ryan Fedasiuk of the American Enterprise Institute said in a LinkedIn post that running a 2.8 trillion-parameter model locally requires hundreds of thousands of dollars of computing equipment.
Trillion Parameter Race
The domestic AI race accelerated recently when Z.ai released GLM-5.2. That model scored near the top US closed-source systems, undermining the consensus that Chinese AI was six months behind.
Parameters are the internal variables a model learns during training and are often used as a rough measure of scale. Direct comparisons remain difficult.
Meituan's LongCat-2.0 and DeepSeek's V4-Pro previously led China with 1.6 trillion-parameters. Anthropic and OpenAI hide parameter counts for their top systems, complicating performance matching.
Moonshot is heavily backed by technology giants Alibaba and Tencent. The startup is seeking $2 billion in fresh funding at a $30 billion valuation ahead of a potential Hong Kong listing, Bloomberg reported last month.
