Qwen3.8-Max Goes Open Weight, With a Licence That Reads Like a Term Sheet
Alibaba has released its first Max-class weights, but the download is not the model on the API and the licence has revenue triggers.
Alibaba has put the weights for Qwen3.8-Max on Hugging Face, the first time any model in the Qwen Max tier has been released for download. The checkpoint went up on 12 August as Qwen/Qwen3.8-2.4T-A95B, in BF16 and FP8 formats, nine days after Alibaba launched the model on its own cloud. It is a mixture-of-experts model with 2.4 trillion total parameters and 95 billion active per token, which puts it just behind Moonshot's Kimi K3 as the largest open-weight model available.
Two days later Alibaba followed with Qwen3.8-27B, a dense model with image and video understanding, under plain Apache 2.0. The contrast between those two releases is the most useful thing to understand about this week.
Same name, different artefact
The downloadable Max is not quite the Max you call through the API. According to Machine Brief's analysis, the open checkpoint is text-only with thinking forced on, while the hosted version keeps multimodal input and a one-million-token context. That is a reasonable engineering choice, since multimodal stacks are harder to package, but it means any benchmark you read from Alibaba's launch post describes a model you cannot fully download. If you are evaluating the open weights, evaluate the open weights.
Alibaba's own numbers are bold, though not uniformly winning. On its internal tests the model scores 86.6 on TerminalBench 2.1, close behind GPT-5.6 Sol's 88.8, and takes the top score of 93 on PaperBench. More eye-catching are the long-horizon case studies The Decoder summarised: a 16-day autonomous software project that produced 265 commits, a five-day research reproduction that launched 33 GPU training jobs, and a simulated online shop that quadrupled its starting capital. These are demonstrations chosen by the vendor, not controlled evaluations, and should be read that way. They do show where the labs think the next contest is: not single answers, but work that runs for days.
The licence is the real news
Qwen built its reputation on permissive licensing. The Max weights break with that. Under the custom licence, a product with 100 million monthly active users or $20 million in monthly revenue must display the model's name prominently. A model-as-a-service or AI work assistant business with more than $50 million in trailing-twelve-month revenue needs a separate paid licence. Internal use that does not expose the model to third parties is exempt.
For most companies those thresholds are far away, and the practical effect is close to free use. But the shape of the terms matters. The paid tier targets exactly the businesses that could turn Alibaba's flagship into a competing hosted service, and they echo the open-until-you-are-big approach Meta used for Llama. Calling this open source would be wrong; open weights with commercial conditions is the accurate description.
The 27B model, by contrast, comes with no strings. Machine Brief reports it scores 61.7 on SWE-Bench Pro. For most teams that want to self-host, that is the more important release this week: a model that fits on hardware people actually own, under a licence lawyers already understand.
What it means
Deployment guidance for the Max points at a full Nvidia GB300 NVL72 rack, which tells you who the weights are really for: cloud providers, governments and research groups. Alibaba gets the reputational benefit of a frontier-scale open release while protecting its hosted business from the largest would-be competitors. That is a sensible strategy, and a sign that the open-weight race at the top end is becoming a commercial negotiation rather than a gift.
Watch for whether other Chinese labs follow with similar revenue-triggered licences, and whether third-party evaluations of the open checkpoint line up with the API model's scores.
Sources