The open weight summer: GLM-5.2, Kimi K3 and Qwen3.8 put frontier scale models in anyone's hands

July: Kimi K3 reaches 2.8 trillion parameters

Beijing based Moonshot AI launched Kimi K3 on 16 July in its apps and API, promising in its launch post that "the full model weights will be released by July 27, 2026". Unite.AI reported the weights on Hugging Face on 27 July. According to the model card, K3 has 2.8 trillion parameters, 104 billion of them active per token, a context window of about 1 million tokens, and it understands images and video as well as text.

Moonshot does not claim the top spot. Its launch post says K3's "overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol". It recommends 64 or more accelerators to run the model, and analysts cited by Reuters expect few customers to run a system of this size on their own hardware because of the computing costs. The release also landed in a political dispute: the White House has accused Moonshot of stealing technology from Anthropic, Chinese officials called the claim unfounded, and Moonshot says its gains came from original changes to its architecture, Reuters reported.

The clause that matters

The Kimi K3 License grants MIT style rights to use, modify, fine-tune and sell the model, with two conditions. A company running a "Model as a Service" (MaaS) business, meaning it gives third parties inference or fine-tuning access with meaningful control over inputs, parameters or training data, must sign a separate agreement with Moonshot before commercial use once it and its affiliates earn more than USD 20 million in any 12 consecutive months. Products with more than 100 million monthly active users or USD 20 million in monthly revenue must display "Kimi K3" prominently. Internal use is exempt.

The threshold counts the revenue of the whole group, not revenue from K3, so every large cloud provider is far above it. The price is not in the licence. Moonshot has asked for a revenue share of up to 30 percent, according to Reuters sources, and on 26 August Reuters reported early stage talks with Microsoft, Amazon and Google about hosting K3 on such terms. The three companies declined to comment; Moonshot did not respond.

August: Qwen3.8 under two licences

Alibaba's Qwen team released its Qwen3.8 weights in mid August, The Decoder reported on 14 August. Qwen3.8-27B is a dense multimodal model with 27 billion parameters under Apache 2.0. Beside it came Qwen3.8-2.4T-A95B, with 2.4 trillion parameters and 95 billion active, which its model card calls the first open release of "a Qwen-Max-class model".

The two do not share a licence. The large model ships under a separate Qwen3.8-Max License: a company running a MaaS business, or an "AI Work Assistant" built mainly for coding or office work, needs a separate licence from Qwen once group revenue passes USD 50 million in any 12 months. Reuters had reported on 7 August that Alibaba planned to ask major users for a share of revenue, at a rate that was still unclear.

DeepSeek stays on MIT, Z.ai moves away

DeepSeek made V4-Pro generally available on 13 August, according to its changelog, and the release checkpoint, DeepSeek-V4-Pro-0813, is on Hugging Face under MIT. The V4 model card lists 1.6 trillion parameters, 49 billion active.

Z.ai went the other way. It launched GLM-5.3, built on the GLM-5.2 base, in mid August but held the weights back for about two weeks of safety testing because of the model's skill at finding and exploiting software vulnerabilities, The Batch reported. OpenAI president Greg Brockman warned on 17 August that open weights models with cyber skills at or near the state of the art would likely "significantly accelerate the threat landscape". When the weights came out, the flagship had left MIT: the GLM-5.3 License requires MaaS operators with more than USD 10 billion in group revenue over 12 months to pass a Z.ai security review, its scope "reasonably determined by Z.AI". The smaller GLM-5.3 Flash was released under MIT.

The fine print, side by side

What it means

For researchers, developers and most companies, every model in the table can be downloaded, fine-tuned and sold as a service without asking anyone: the conditions start at revenue levels few firms reach. They target clouds and large software companies and count group revenue, not what the model earns. For those buyers, the licence file now says more than the parameter count.

In the United States, Lin Qiao, chief executive of Fireworks AI, sees no "fundamental barrier" to powerful open models from American labs. "We are really waiting for that to happen," Qiao told Reuters.