Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter model that the company says is the largest open-source AI model in the world, VentureBeat reported. The same day, K3 reached first place on Arena.ai's Frontend Code leaderboard with 1,679 points, ahead of Anthropic's Claude Fable 5 (1,631) and OpenAI's GPT-5.6 Sol (1,618), according to Notebookcheck. It was the first Chinese model to hold that spot. Its predecessor had ranked 18th on the same board.
The Frontend Code board is one leaderboard, based on human preference on a coding subset. Notebookcheck calls it "just one reference point." On the broader Artificial Analysis Intelligence Index, K3 debuted fourth overall, behind Fable 5 and GPT-5.6 Sol and ahead of Claude Opus 4.8 and SpaceX's Grok 4.5, according to an August 3 Forbes contributor column by Drew Bernstein, co-chairman of Marcum Asia. Moonshot itself said K3 outperforms all rivals except Claude Fable 5 and GPT-5.6 on overall capability, as Bloomberg reported in a story carried by The Straits Times.
Moonshot published the full weights on July 27, together with a technical report (Kimi announcement). They ship under a custom Kimi K3 License, not a standard open-source license. Per the license on Hugging Face, a company that runs a "Model as a Service" business with more than $20 million in revenue over 12 months must sign a separate agreement with Moonshot before commercial use.

The distillation question
Bernstein wrote that a screenshot made the rounds in late July of a user asking K3 to identify itself, and the model replying that it was "Claude, an AI assistant" made by Anthropic.
In February, Anthropic published a report on what it called "industrial-scale campaigns" by three labs, DeepSeek, Moonshot and MiniMax, to extract Claude's capabilities. Anthropic says the labs generated over 16 million exchanges with Claude through about 24,000 fraudulent accounts. It puts Moonshot's campaign at over 3.4 million exchanges, and says that in a later phase Moonshot tried to extract and reconstruct Claude's reasoning traces.
According to Bernstein, White House science adviser Michael Kratsios accused Moonshot on July 22 of building an internal platform to distill Anthropic's Fable model for K3. Treasury Secretary Scott Bessent has said he might add Moonshot to a trade blacklist, and Moonshot has disputed claims that K3's performance was achieved through distillation, Reuters reported on September 3.
Moonshot points to its architecture. The K3 model card says the model is built on Kimi Delta Attention and Attention Residuals, and uses a Stable LatentMoE framework that "activates 16 out of 896 experts." Bernstein wrote that skeptics counter that if K3's training were that efficient, its inference costs should be much lower than those of US peers. At launch, Notebookcheck listed K3's API price at $3 per million input tokens and $15 per million output tokens, the same standard rate as Anthropic's Claude Sonnet 5.
US companies already use Chinese open models
Bernstein cites these signs that US companies use Chinese open models:
- Airbnb CEO Brian Chesky said last fall that his company relies heavily on Alibaba's Qwen models.
- Andreessen Horowitz partner Martin Casado has estimated an 80% chance that any given startup pitching his firm is building on a Chinese open-source model.
- Chinese-origin models account for nearly half the tokens routed through OpenRouter, up from roughly 11% a year earlier.
His explanation is cost: "open weights let companies fine-tune and self-host at a fraction of U.S. frontier prices, with no API contract and no data leaving their infrastructure." For open-weight models you can run today, see our guide to the best open-source LLMs.
A Hong Kong listing
On July 19, Bloomberg reported that Moonshot had distributed a shareholder resolution seeking its investors' approval for a Hong Kong listing, which could happen as soon as within six months. Per the same report, as carried by The Straits Times, Moonshot was wrapping up a funding round that may value it at more than $30 billion, and its annual recurring revenue hit $300 million in June. Moonshot has begun dismantling its "red chip" structure, which the report calls "a requirement for a smoother overseas listing process."
On September 3, Reuters reported, citing three people with knowledge of the plans, that Moonshot had confidentially filed for a Hong Kong IPO and aimed to raise $3 billion. Two sources told Reuters that Moonshot has been valued at $50 billion in an ongoing funding round. Reuters also cited earlier reports from sources that Moonshot is in talks with Microsoft, Amazon and Google on revenue-sharing deals to host K3, which we covered in August. Z.AI, formerly Zhipu, and MiniMax listed in Hong Kong this year, Reuters noted.
Bernstein argues that the listing will let public markets "referee the U.S.-China AI race with audited numbers."
Open and closed, as of early August
Bernstein also argues that Chinese labs are closing some models as they succeed, while some US labs are opening up. As of his August 3 column, he listed:
- MiniMax kept its latest model closed after two open flagship generations.
- Zhipu released its newest GLM flagship as proprietary.
- Alibaba keeps the Qwen family open while its best Max-tier models stay behind an API.
- OpenAI shipped gpt-oss last year, its first open-weight release in six years, and Meta released its newest flagship closed.
- Nvidia released its 550-billion-parameter Nemotron 3 Ultra under an open license in June, and Thinking Machines Lab launched Inkling, a 975-billion-parameter open-weight model, on July 15.
His conclusion: "challengers open up to buy distribution, and leaders lock down to harvest it."
Update, October 26, 2026
Moonshot released Kimi K3.1 Pro on October 23, TSN reported. The launch blog calls it "a post-trained version of Kimi K3, with substantially larger post-training compute," with a 1-million-token context window, and says "The full model weights will be released by Nov 4, 2026." TSN reports the K3.1 Pro API price of $2 per million input tokens and $10 per million output tokens is below the $3 and $15 it lists for K3. On October 25, Moonshot apologized for rate limits on the K3.1 Pro subscription plan and reset 25% of every subscriber's monthly quota. Reuters, in a report carried by CNA, noted that Moonshot also had computing capacity problems after K3 launched. Our Kimi K3.1 Pro explainer covers pricing, benchmarks and the quota reset.
Update, November 4, 2026: Moonshot has since published the weights on Hugging Face (moonshotai/Kimi-K3.1-Pro, created November 1, 2026, Pacific time). The October 26 update above gives the date Moonshot promised.
Sources
- Forbes contributor column (Drew Bernstein): Chinese AI Models At The Frontier, August 3, 2026
- VentureBeat: China's Moonshot AI releases Kimi K3, July 16, 2026
- Notebookcheck: Kimi K3 tops AI benchmark in a first for Chinese models
- Kimi: Kimi K3 open day, weights and technical report (July 27, 2026)
- Kimi K3 model card (Hugging Face)
- Kimi K3 License (Hugging Face)
- Anthropic: Detecting and preventing distillation attacks
- Bloomberg via The Straits Times: China's Moonshot plans IPO in 6 months after AI breakthrough
- Reuters via MarketScreener: Moonshot files confidentially for Hong Kong IPO, September 3, 2026
- Kimi K3.1 Pro launch blog (Moonshot AI)
- TSN: Kimi K3.1 Pro launches, weights due by 4 Nov
- CNA: Moonshot resets subscriber quotas after Kimi K3.1 Pro launch issues
