The weights are up. Alibaba published Qwen3.8-2.4T-A95B on Hugging Face on Friday, and the model card makes the claim directly: “For the first time, Qwen3.8 brings a Qwen-Max-class model to open release.” 2.4 trillion total parameters, 95 billion activated per forward pass, 262,144 tokens of native context extensible to roughly 1.01 million.
We have now covered this model twice without being able to answer the only question that mattered. When the preview appeared in July, the weights were promised “soon” with no licence and no benchmarks. When Qwen3.8-Max launched on August 3, the benchmarks arrived and the weights were still pending; we flagged the open item as “the weights, and the license — whether the license is Apache-clean or commercially restricted decides how real the open-weight threat is.”
It is not Apache-clean.
What the licence actually says
The repository metadata lists the licence as qwen3.8-max, and the repo ships a LICENSE file of that name rather than a standard one. The reported terms: an MIT-style grant, with an attribution requirement that switches on once a deployment crosses 100 million monthly active users or $20 million in monthly revenue, and a separate paid licence required for model-as-a-service or AI work-assistant products above $50 million in aggregate trailing-twelve-month revenue.
For the overwhelming majority of companies that is functionally free. For the specific set of companies that would resell this model as a product — the ones whose competitive pressure would actually reprice the market — it is a negotiation with Alibaba. The tripwire is aimed precisely at the businesses the open release was supposed to threaten.
The second catch: open is not the same product as hosted
Per the model card, the downloadable model is text-only. Thinking mode cannot be disabled. Native context is 262K, not the advertised million. The hosted Qwen3.8-Max adds image and video input, optional thinking, a 1M default context and built-in tools — and stays on Alibaba Cloud. The most-engaged thread on the repository’s discussion board is a complaint about exactly this gap.
That matters because the multimodality was the differentiator. Alibaba’s July pitch was its first multimodal model above a trillion parameters, processing images, video and documents. The open release does none of that.
One genuinely permissive release did ship: Qwen3.8-27B, published two days earlier, is a dense multimodal model under a real Apache 2.0 licence with the same 262K native context. The small one is open. The big one is conditional.
Our take: This is not a broken promise — downloadable frontier-class weights are a real event and the walk-away leverage in enterprise contract talks is real. But it is a narrower event than the headlines suggest, and the shape is worth naming: open weights are converging on customer acquisition with a revenue trigger, not donation. Compare it to Meta’s Muse Glimmer release last week, which is Apache 2.0 with no threshold at all — a materially different offer, at a much smaller scale. The practical test for any team evaluating this: if your product is an AI assistant and you are anywhere near $50 million in trailing revenue, read the LICENSE file before the model card. If you need vision, the open weights do not do the job at any price.
What to watch
- Whether revenue thresholds become the norm. If the major Chinese labs converge on tripwire licences, “open weights” stops being a binary and starts being a term sheet.
- Moonshot’s response. The July release was aimed at Kimi K3’s IPO run. A licence with a commercial carve-out blunts that weapon somewhat.
- Serving economics. 2.4T parameters in BF16 is a datacentre commitment. Community FP8 and quantised builds already exist; how far the footprint falls decides who can actually run it.
- Meta’s Muse Spark 1.2 weights. Still promised, still unshipped. A US frontier model under a permissive licence would reset the whole comparison.
- Whether the hosted-versus-open feature gap widens. Vision, tools and 1M context stayed behind the API this round. Watch the next one.