This blog has tracked Qwen3.8-Max since it was a bare parameter count with no model card. On August 12, the last piece landed: weights for Qwen/Qwen3.8-2.4T-A95B and its FP8 variant went live on Hugging Face, confirmed directly against both repos. Alibaba is calling it a milestone — “for the first time, Qwen3.8 brings a Qwen-Max-class model to open release,” per the model card. That framing is accurate, narrowly. It’s also doing a lot of work to obscure how differently “open” behaves here than the last two Qwen generations.
The timeline says more than the announcement#
Alibaba previewed Qwen3.8-Max at WAIC Shanghai on July 19 with a parameter count and nothing else. API access went GA on August 3, paired with a promise that open weights were coming “next week.” August 10 came and went with no Hugging Face repo — trackers watching the org page found nothing. The weights actually appeared August 11-12, roughly three weeks after “next week” was first said out loud.
Three weeks isn’t a scandal. But it’s the second time in this model’s rollout that “open” arrived later than announced, and the gap is worth naming plainly rather than letting the eventual release wash it out. A New Stack piece on developer reaction to the delay summed up the mood as watching “an API business model wearing an open source jacket” — harsh, but it’s describing the sequencing accurately: paid API first, open weights as an afterthought delivered under pressure.
The license: not new, but a real reversal#
Here’s the part most coverage is getting slightly wrong. The Qwen3.8-Max license — labeled qwen3.8-max in the repo, not Apache 2.0 — has two operative conditions, confirmed directly from the LICENSE file:
- If your product or service crosses 100 million monthly active users or $20 million in monthly revenue, you have to display the model name prominently in your UI.
- If you’re running a Model-as-a-Service or “AI Work Assistant” business (Alibaba’s terms for inference-hosting or standalone coding/office AI products) generating more than $50 million in aggregate revenue over any 12-month period, you need a separate commercial license from Qwen before you can keep using it for that purpose. Pure internal use, with no third-party access, is exempt.
Several outlets have headlined this as a “revenue-share” license, drawing a direct line to Moonshot’s Kimi K3 terms (reported at a 30% cut above $20M). That comparison doesn’t hold up against the primary text — there’s no percentage anywhere in the Qwen3.8-Max LICENSE file, just a requirement to go negotiate a separate license once you cross the threshold. It’s a usage-gate, not a revenue-share. Worth correcting, because the distinction matters if you’re actually trying to figure out what building on this model would cost you.
What’s genuinely notable isn’t that Alibaba wrote a custom license — it’s that this is a reversal, not a first. Qwen-72B and Qwen2-72B both shipped under near-identical custom terms years ago, complete with the same 100-million-MAU commercial trigger (Qwen2’s license went further, with a non-compete clause barring you from using outputs to improve any non-Qwen model). Then Qwen3 — the entire generation, flagship included — went fully Apache 2.0, no exceptions, the most permissive release Alibaba had shipped. Qwen3.8-Max walks that back. The real, narrower “first” here is that no prior Qwen-Max-tier model — the actual flagship class, as opposed to “biggest open-weighted size” — had ever been open-weighted at all; Qwen2.5-Max and Qwen3.5-Max both stayed API-only. Alibaba is trading permissiveness for reach on its best model, in exchange for a licensing hook it can pull if you get big enough to matter.
The benchmarks are Alibaba’s, and only Alibaba’s#
The model card’s self-reported numbers are competitive on paper: 86.6 on Terminal-Bench 2.1, 67.7 on SWE-bench Pro, 93.0 on PaperBench, 92.6 on GPQA Diamond, compared directly against Claude Opus 4.8, Claude Fable 5, and GPT-5.6 Sol in Alibaba’s own table.
None of it is independently verified. Scale AI’s public SWE-bench Pro leaderboard has no Qwen3.8-Max entry as of this writing, and no Artificial Analysis or LMArena score exists either. That’s not an accusation of fraud — it’s the same evidentiary gap this blog flagged on xAI’s self-reported Grok 4.5 SWE-bench Pro number a week ago, and it deserves the same treatment here: a lab’s own benchmark table is a claim, not a result, until someone without a stake in the outcome reproduces it. Readers should also watch for search-engine confusion between “Qwen3.8” and the real, pre-existing “Qwen3-8B” (an unrelated 8-billion-parameter dense model) — the naming collision is already generating mixed-up coverage.
“Open” doesn’t mean “runnable”#
Even setting the license aside, actually self-hosting the flagship checkpoint is out of reach for the overwhelming majority of the people this announcement is nominally aimed at. Unsloth’s own quantization guide puts the cheapest usable configuration — a dynamic 1-bit quant — at roughly 450GB of combined RAM and VRAM, with the lossless BF16 version running to 4.9TB. Alibaba’s own reference deployment, published via NVIDIA, uses 72 Blackwell Ultra GPUs in a single NVLink domain. This is not a model you run on a workstation; it’s a model you rent from a cloud provider, which somewhat defeats the practical case for “open weights” as an alternative to an API in the first place.
The smaller sibling shipped alongside it, Qwen3.8-27B, is the actually-accessible release — pitched at 16GB+ VRAM setups — and gets a fraction of the attention the 2.4T flagship is generating. If you care about running open weights on hardware you own, that’s the model worth looking at, not the one making headlines.
Why this matters beyond one release#
None of this makes Qwen3.8-Max uninteresting — a 2.4-trillion-parameter MoE with a 1-million-token extended context is a real engineering achievement, and Alibaba shipping it in any open form is more than most labs at this scale do for their actual flagships. But “open weights” has become a marketing category that papers over enormous variation in what “open” costs you: license terms that gate you the moment you succeed, benchmark claims nobody outside the lab has reproduced, and hardware requirements that put “self-hosting” out of reach for anyone without a data center. Claude Code doesn’t claim to be open-weight and never has — Anthropic’s argument has always been that you’re paying for a hosted, continuously-updated agent, not a checkpoint you own. Say what you will about that trade, at least it’s the trade being advertised. Qwen3.8-Max is being sold as something closer to open than the license, the hardware floor, or the unverified benchmark table actually support.
Sources: Qwen/Qwen3.8-2.4T-A95B model card and LICENSE (Hugging Face, fetched directly), Qwen-72B and Qwen2-72B license text for historical comparison, Unsloth Qwen3.8 deployment guide, NVIDIA developer blog on GB300 NVL72 deployment, Scale AI’s public SWE-bench Pro leaderboard (checked for independent verification, none found).