Kimi K3 Becomes the World's Largest Open Weight AI Model as Moonshot Releases It for Free

Moonshot AI releases Kimi K3's weights

Moonshot AI published the open weights for Kimi K3 on Monday, July 27, 2026, a day ahead of the deadline the company itself had been communicating. According to TechTimes, Qz and StartupFortune, the release went live around 7:30 p.m. Eastern Time on Sunday, July 26, making available to the public the largest open weight model ever released. The model is published on Hugging Face under a Modified MIT license, allowing anyone to download, inspect, fine tune and run the system on their own infrastructure.

A giant model even in compressed form

Kimi K3 has 2.8 trillion parameters arranged in a sparse mixture of experts architecture, a 1,048,576 token context window and native support for text, images and video. Even at reduced 4 bit MXFP4 precision, the weights take up roughly 1.4 terabytes of fast memory, versus about 5.6 terabytes at 16 bit, according to TECHi. Loading the model therefore requires a cluster of top tier GPUs, out of reach for the vast majority of companies.

Day zero access through Together AI and Modal

For anyone without that kind of infrastructure, Together AI and Modal announced hosted access on launch day itself, according to explainx.ai, letting developers use Kimi K3 through an API without downloading 1.4 terabytes of weights or running their own hardware. That kind of intermediary is, in practice, what will bring the model to most interested developers and companies.

Why open weights differ from API access

The difference between an open weight model and a closed model accessed only through an API is not subtle: with the weights in hand, any organization can audit the model's behavior, fine tune it with its own data and run it entirely within its own infrastructure, without depending on a vendor's uptime or data policy. TechTimes highlights exactly this point, noting that hosting Kimi K3 locally reduces the risk of sensitive data passing through servers outside the controlling company's reach, an argument that carries extra weight for a Chinese origin model used by Western companies.

One more rung in the open model race

Today's release is the latest chapter in a story this outlet has been tracking: Kimi K3 originally arrived in mid July as the world's most advanced open model, then had to pause new subscriptions due to overwhelming demand, prompted xAI's Grok 4.6 as a direct response, and saw independent tests report a 51 percent hallucination rate, a figure absent from Moonshot's own published benchmarks. Today's open weight release is a different kind of step: it is no longer about subscribing to access the model, it is about owning the model.

Political tension in Washington

The release lands amid a heated US debate over Chinese open models, after American government officials publicly accused Moonshot of distilling Kimi K3 from Anthropic's Claude Fable 5, a claim Moonshot denies. That political tension is the direct backdrop to a letter signed by 25 American technology companies, including Nvidia and Microsoft, asking the US government not to restrict open weight models, a story this outlet covers separately.

Why it matters to Brazilian agencies and SMBs

Running Kimi K3 on your own infrastructure is out of the question for practically any Brazilian agency or SMB, given the model's size. But the indirect effect matters: the more frontier open weight options exist, hosted by vendors like Together AI and Modal at competitive prices, the more pricing pressure builds on closed models from OpenAI, Google and Anthropic. For anyone buying AI model access through an API for customer service and marketing automation, every new heavyweight open model that reaches the market is, in practice, one more bargaining chip when picking a vendor and negotiating price.