Moonshot AI

“A Pivotal Moment”: Kimi K3 Open-Weight Release Just Hours Away

co-written by eustella AI27. July 2026, 13:49

Moonshot AI is releasing the weights of its Kimi K3 model for free download this afternoon. Interest is considerable: on the official model page at Hugging Face, around 2,500 developers have registered to be notified as soon as the download goes live. At 2.8 trillion parameters, Kimi K3 is the largest AI model ever released publicly.

The Beijing lab behind the Kimi assistant had already launched K3 via API and web app on 16 July, but held back the weights for ten days. Now they are public – roughly 594 gigabytes in native MXFP4 format, and up to 1.4 terabytes depending on quantisation. One distinction matters here: “open weight” is not “open source”. What is published are the trained weights (expected under a modified MIT licence), not the training data or the training pipeline.

Technically, K3 is a sparse mixture-of-experts model with a context window of one million tokens, native vision capabilities and a focus on long, autonomous agent runs. Artificial Analysis ranks it third on its Intelligence Index – behind Claude Fable 5 and GPT-5.6 Sol Max, at a considerably lower price. In the Frontend Code Arena, K3 sits in first place. The API costs $3 per million input and $15 per million output tokens, roughly half of Anthropic’s Opus 4.8. Its weak point is generally seen as speed.

The Bottleneck Is Memory

Anyone wanting to run the model themselves runs into a physical limit: Moonshot recommends supernodes with 64 or more accelerators. Eugene Cheah, CEO and co-founder of the inference platform Featherless.ai, sees this as the real hurdle of the release:

“The release of Kimi K3 from Moonshot AI represents a pivotal moment in the international AI ecosystem as it proves that open-weight AI models are capable of competing directly with proprietary frontier models such as Fable 5. Nonetheless, running a model with 2.8 trillion parameters requires a new level of hardware capabilities.”

The problem, he argues, is less compute than memory: “The NVIDIA B200 devices’ lower capacity makes them more difficult and costly to perform calculations of this level, and so, companies have to seek B300 processors, which are nearly sold out at the moment.”

Featherless – which closed a $20 million round led by AMD Ventures and Airbus Ventures in April, and therefore has a vested interest in this diagnosis – is betting on AMD hardware: “Instead of having to pay more than $50,000 per month for B300 clusters, we enable engineering teams to use K3 from $15,000 per month.” On top of that comes the sovereignty argument: “By using K3 on private cloud infrastructure in the EU and US with zero-logging policies, we provide organisations with complete sovereignty over their AI pipeline.”

The Bigger Fight in the Background

The release lands in the middle of an open confrontation in Washington. After K3 was announced, chip stocks fell on the Nasdaq and the Nikkei, and observers drew parallels to the “DeepSeek moment” of 2025. On 22 July, White House officials accused Moonshot AI of siphoning off American intellectual property through distillation; Anthropic’s policy chief spoke of industrial espionage, and Treasury Secretary Scott Bessent floated sanctions. So far there is no publicly verifiable evidence for the allegation, and no ban has been drafted.

The counterstrike came within 48 hours. Around 200 startups wrote to President Trump, asking him not to take away the open models they build on. A day later, Jensen Huang published the first post of his life on X – sharing a three-page letter titled “Open Weights and American AI Leadership”, which warns policymakers against “premature restrictions” on downloadable models. It was initially signed by 25 companies and organisations, among them Nvidia, Microsoft, Meta, IBM, Dell, Palantir, ServiceNow, CrowdStrike, Perplexity, Replit, Mistral, Hugging Face, Mozilla, Andreessen Horowitz, Y Combinator and the Linux Foundation. Within a day the list doubled to 50 names, now including OpenAI, Google, AMD, Cisco, Cloudflare, GitHub and Ollama. Anthropic and Amazon have not signed.

A Game of Weights

What stands out is the hypocrisy on both sides. Nvidia organised the letter while controlling CUDA, the most proprietary software layer in the industry. Microsoft opens no operating systems. Meta’s Llama licence contains usage restrictions and, according to an analysis from circles close to the EU AI Office, does not meet the AI Act’s open-source standard. Openness costs these companies nothing where they demand it. Conversely, investor Bill Gurley countered Anthropic by noting that its safety concerns happen to arise precisely where open models attack its own business model. Both readings can be true at once – an honest safety position and a tangible commercial interest are not mutually exclusive.

Ultimately, this is less about ideology than about business models: whoever sells hardware, clouds, tools and applications benefits from every additional deployment. Whoever sells intelligence per token loses pricing power the moment a Chinese lab gives away a frontier-level model. The fact that Beijing’s Ministry of Commerce is simultaneously weighing export controls on model weights and training data, according to the Financial Times, shows how far both superpowers have come to treat open weights as a strategic asset. For European companies that have built their AI strategies on the availability of open models, an assumption that until recently seemed self-evident is now in question.