28 July 2026

^{Beijing skyline (photo: Albert Canite on Unsplash)}

 

Moonshot AI has released the full weights of Kimi K3, giving developers access to one of the largest open-weight artificial intelligence models yet built.

The Beijing-based AI startup describes Kimi K3 as a 2.8-trillion-parameter, natively multimodal model designed for coding, knowledge work and complex reasoning. It can process up to one million tokens in a single context, accepts text and images, and is also presented by Moonshot as capable of understanding video.

Moonshot acknowledges that Kimi K3 still trails the most powerful proprietary AI models overall, but says it performs competitively with leading systems from OpenAI and Anthropic across several evaluations.

The release brings those claims into the open. Developers can now download the weights, examine the model and adapt it for their own applications under Moonshot’s Kimi K3 licence.

Kimi K3 is open-weight rather than fully open-source. Its model weights and accompanying code are available, but Moonshot has not disclosed all the data and processes used to train it. That means outside researchers can inspect and modify the finished model without being able to reproduce its entire development.

The licence also contains commercial conditions. Companies operating K3 as a model service must reach a separate agreement with Moonshot if their total annual revenue exceeds $20 million. Products using the model must prominently display the Kimi K3 name if they exceed 100 million monthly active users or $20 million in monthly revenue.

K3 is also far from easy to operate. Although its Mixture-of-Experts architecture selects only 16 of 896 experts for each token, 104 billion parameters are active during inference. The complete model remains enormous. Moonshot recommends deployments using “supernode” configurations with at least 64 accelerators.

Open access to Kimi K3 therefore does not place frontier AI within reach of an ordinary developer with a workstation. It does, however, give cloud providers, research institutions and well-resourced companies an alternative to proprietary models available exclusively through an application programming interface.

The release also brings Moonshot (and the controversy surrounding its technological progress) into the global developer ecosystem.

 

Who founded Moonshot AI?

 

Moonshot AI was founded in Beijing in 2023 by Yang Zhilin, Zhou Xinyu and Wu Yuxin, all alumni of Tsinghua University. Its Chinese name, Yuezhi Anmian, translates as “the dark side of the moon”, reflecting Yang’s enthusiasm for Pink Floyd and the band’s album of the same name.

The musical reference can make the company sound more eccentric than it is. Moonshot’s founders brought substantial AI research and engineering experience.

Yang studied computer science at Tsinghua before completing a doctorate at Carnegie Mellon University. During his academic career, he worked at Google Brain and Meta and studied under prominent machine-learning researchers including Ruslan Salakhutdinov and William Cohen.

Before founding Moonshot, Yang co-authored two influential language-modelling papers. The first introduced Transformer-XL, an architecture designed to help AI models preserve information across longer sequences instead of treating each segment largely in isolation.

The second introduced XLNet, an alternative language-pretraining method that outperformed Google’s BERT across 20 tasks in the researchers’ original evaluation.

Both papers appeared in 2019, years before ChatGPT turned large language models into consumer products. They provide important context for Moonshot’s later emphasis on model memory and long-context processing. The company did not adopt long context simply because it became a fashionable benchmark; the problem had been central to its founder’s research career.

Wu also brought experience from leading American AI laboratories. A Tsinghua and Carnegie Mellon graduate, he worked on foundation models at Google Brain and computer vision at Meta AI Research. He created Detectron2, Meta’s widely used software platform for object detection, segmentation and other visual-recognition tasks.

That background has become increasingly relevant as Moonshot has moved beyond text towards multimodal AI models capable of interpreting images and visual interfaces.

 

How Kimi evolved from document reader to frontier model

 

Moonshot introduced its Kimi chatbot in October 2023. Its initial differentiator was the ability to process as many as 200,000 Chinese characters in one conversation. That gave Kimi practical applications in analysing research papers, legal documents, financial reports and books.

Moonshot later increased the supported context and expanded Kimi into coding, web research and agent-based work. In July 2025, it released the weights of Kimi K2, a one-trillion-parameter Mixture-of-Experts model. Subsequent versions added stronger reasoning, visual understanding and software-development capabilities.

Kimi K3 takes that progression considerably further. Its 2.8 trillion parameters are distributed across hundreds of specialised components, although 104 billion parameters are active during each inference step.

Moonshot says two new mechanisms—Kimi Delta Attention and Attention Residuals—improve how information passes across long sequences and through the model’s layers. The company claims these architectural changes, greater sparsity and revised training methods have made scaling approximately 2.5 times more efficient than with Kimi K2.

 

Who has invested in Moonshot AI?

 

Moonshot has attracted considerable financial backing. Reuters reported that the company raised more than $2 billion in May 2026 from investors including Meituan, China Mobile and CPE.

According to a fundraising document seen by the news agency, the round took Moonshot’s historical fundraising above $5.5 billion. Earlier investors included Alibaba and Tencent.

Reuters also reported that Moonshot was seeking as much as another $2 billion after reaching a valuation of $30 billion in June. Advisers including Goldman Sachs and China International Capital Corp were discussing a possible Hong Kong initial public offering, although Moonshot declined to comment on the reported listing preparations.

The cost of its ambitions is already visible. Moonshot temporarily paused new consumer subscriptions after demand for Kimi K3 strained its computing capacity. The interruption demonstrated both the interest in the model and the infrastructure burden created by serving it.

 

Why Anthropic has accused Moonshot of model distillation

 

Moonshot’s rapid progress has attracted scrutiny from Anthropic and the US government.

In February, Anthropic said it had identified more than 3.4 million interactions with its Claude AI models that were linked to Moonshot. According to Anthropic, hundreds of fraudulent accounts were used to target capabilities including reasoning, coding, data analysis, computer vision and tool use.

Such interactions can be used for model distillation: training or improving one AI system with outputs generated by another. Distillation is a widely used machine-learning technique and is not inherently unlawful.

The dispute concerns the scale and method of access, whether platform restrictions were deliberately circumvented and whether the activity constituted the extraction of proprietary AI capabilities.

Following Kimi K3’s launch, White House Office of Science and Technology Policy director Michael Kratsios said the US government had information indicating that Moonshot used Anthropic’s Claude Fable 5 in developing K3.

Kratsios alleged that Moonshot had built an internal platform to conduct large-scale distillation of US models and switched between different access methods to avoid detection.

 

The Nvidia GB300 export-control allegations

 

Kratsios separately alleged that Moonshot had acquired servers fitted with Nvidia GB300 processors and accessed similar systems in Thailand, potentially circumventing US restrictions on supplying advanced Blackwell AI chips to Chinese companies.

The US government has not publicly released evidence supporting those allegations.

Moonshot has rejected suggestions that Kimi K3’s performance resulted from distillation, telling China’s National Business Daily that its gains came from original changes to the model’s underlying architecture.

China’s commerce ministry has also entered the dispute, accusing Washington of threatening companies on the basis of allegations it said lacked factual and legal foundations. The ministry said China would take “all necessary measures” if US actions caused substantive harm to Chinese interests.

US Treasury Secretary Scott Bessent, meanwhile, has warned that Chinese companies could face sanctions or placement on the Commerce Department’s Entity List if the US concludes that industrial-scale distillation crossed into intellectual-property theft. Moonshot had not been added to the list at the time of writing.

 

What Kimi K3’s open weights cannot reveal

 

Releasing Kimi K3’s weights will allow researchers to test Moonshot’s performance claims, study its architecture and observe how developers use it. It cannot establish precisely which data, model outputs or computing systems contributed to its training.

Moonshot’s own record complicates any simple explanation of its rise. This is not a company without original research credentials: its founders have worked on long-context modelling, computer vision and AI infrastructure since before the current generative-AI boom.

Nor can the accusations be dismissed solely as a predictable response to Chinese competition. Anthropic and senior US officials have made specific claims about account activity, distillation infrastructure and hardware access, although the evidence behind several of those claims remains undisclosed.

Kimi K3 consequently arrives as more than another AI model release. It is a test of whether open weights can reshape the global AI market, but also a case study in how difficult it has become to separate scientific progress, commercial competition and national industrial policy.

The model can now be downloaded. The argument over how Moonshot built it will be much harder to resolve.

 

Further reading on MoveTheNeedle. news:

Can OXMIQ’s OxPython turn CUDA against Nvidia?

Europe's next strategic dependency may be AI itself

Is Claude becoming the AI equivalent of Windows, SAP or AWS?