Qwen3.8-27B Is Apache 2.0. The 2.4T Max Weights Have No Published Licence Text
A.I. / news
Qwen3.8-27B Is Apache 2.0. The 2.4T Max Weights Have No Published Licence Text
eWeek reports a $50 million revenue trigger on the Max licence; RuntimeWire calls the terms unresolved, and the Hugging Face card gives only a name.
Alibaba's Qwen3.8-27B is listed as Apache 2.0 on its Hugging Face card, while the 2.4-trillion-parameter Qwen3.8 Max weights released alongside it carry a licence the card names qwen3.8-max and does not reproduce. Two outlets disagree on what that second licence says.
Weeks after release, the 27B model still sits on Hugging Face's trending list with 17,407 likes and 6,768,654 downloads as of Saturday. The licence split is the part of the release that decides who can build a product on it.
What the 27B card says
The Qwen3.8-27B card lists Apache 2.0, a dense model with a vision encoder, a native 262,144-token context and extension to 1,000,000 tokens. It says the model is "built on the architectural foundation of Qwen3.5". Weights are BF16, there is no access gate, and the card gives no hardware guidance.
Release timing is loose. eWeek said the weights went up "on Friday" and dated its article August 17, 2026. The collection page reads "Updated Aug 13".
The licence on the big model
The Qwen3.8-2.4T-A95B card lists 2.4 trillion total and 95 billion active parameters, 512 experts with 10 routed and one shared per token, and a text-only design where "thinking cannot be disabled". Its licence field reads "qwen3.8-max". The page has no licence text, no revenue threshold and no list of restrictions.
eWeek reported that businesses running a qualifying "model as a service" or "AI work assistant" need a separate licence above $50 million in revenue over any consecutive 12 months, with internal use exempt. RuntimeWire reviewed the same release and wrote that the Max licence "remains unresolved", citing no official Alibaba document setting terms.
The $50 million figure appears in one source, so treat it as unconfirmed until Alibaba publishes the licence text.
Benchmarks, vendor-supplied
All figures below come from Qwen's own cards. The 27B card lists Claude Opus 4.6 Max ahead on some text benchmarks, among them Terminal Bench 2.1 at 78.2 and GPQA Diamond at 91.3.
| Benchmark | Qwen3.8-27B | Qwen3.8 Max |
|---|---|---|
| Terminal Bench 2.1 | 73.0 | 86.6 |
| SWE-bench Pro | 61.7 | 67.7 |
| GPQA Diamond | 89.2 | 92.6 |
| HLE | 30.8 | 43.6 |
- Qwen3.8 Max86.6 points
- Claude Opus 4.6 Max78.2 points
- Qwen3.8-27B73 points
Source: Qwen3.8-27B and Qwen3.8-2.4T-A95B model cards, accessed 2026-10-10
eWeek also relayed MLQ.ai's point that the release materials lack a full like-for-like comparison with earlier Qwen models and competitors on identical prompts and evaluation setups. The two Opus figures are the only cross-vendor numbers quoted here.
What it means for builders
The 27B model is the one with clear terms. It is permissive enough for a commercial product, subject to the Apache 2.0 conditions, in eWeek's reading, and Microsoft's Decision-1 and Cloudflare's Clef both build on Qwen bases.
Anyone planning on the 2.4T weights has no published licence text to build on. Other Apache 2.0 releases on this site include Google's EmbeddingGemma 2.
None of the sources reviewed says when the Max licence text will be published.
Sources
More in A.I.
- 01Nvidia Is Reportedly Weighing a Takeover of Reflection AI, Whose 501B Beam Model Has No Public Weights YetThe Financial Times reported early-stage talks on October 10. Reflection's own benchmark table has Beam behind Kimi K3 and Qwen 3.8 Max on every coding test where all three report.
- 02OpenAI Posts Hundreds of AI-Written Maths Papers, Keeps the PromptsThe repository holds 719 manuscripts by README count, a 42% Lean share by OpenAI's measure and 22% by Decrypt's, and no model name.
- 03Cloudflare Open-Sources Clef, a 27B Drop-In for TypeSafe's JevThe Apache 2.0 decision models cut median latency from 524 ms to 39 ms on Cloudflare's own tests, and lose to Jev by 30 points on GPQA Diamond.
- 04Microsoft-Decision-1 Is a Post-Trained Qwen3.5-9B at $0.042 per Million Tokens, With No Licence GivenMicrosoft's launch post claims a win across 36 benchmarks and 35 times the speed of GPT-6 Sol, and every figure in it is the company's own.