Qwen 3.8-Max Ships Open Weights  -  but the Fine Print Isn't Apache

Alibaba published the first Max-class Qwen model as downloadable weights on August 12, but under a custom Qwen3.8-Max License with revenue thresholds, not Apache 2.0. The text-only checkpoint, Qwen3.8-2.4T-A95B, is available on Hugging Face in BF16 and FP8 variants, while the hosted API retains multimodal input and 1M-token context. The license requires prominent model name display for products exceeding 100 million monthly active users or $20 million monthly revenue, and a separate paid license for model-as-a-service or AI work assistant businesses with over $50 million trailing-twelve-month revenue. In contrast, the sibling Qwen3.8-27B, released August 14 under Apache 2.0, scored 61.7 on SWE-Bench Pro.

Qwen 3.8-Max Ships Open Weights - but the Fine Print Isn't Apache Alibaba published the first Max-class Qwen open weights, but under a custom license with revenue thresholds, not Apache 2.0. The downloadable checkpoint… Alibaba made history this week by publishing the first Max-class Qwen model as downloadable weights. Then it attached a license that quietly changes what open means in the AI industry. Qwen 3.8-Max weights shipped August 12 on Hugging Face /glossary/hugging-face under the repo Qwen/Qwen3.8-2.4T-A95B, in BF16 safetensors plus an FP8 variant. Nvidia /glossary/nvidia 's own deployment engineering blog confirmed the release the same day while walking through how to serve it on a GB300 NVL72 rack. This is the first time a Max-class Qwen model has been downloadable, and it is not the model the API sells. What Got Stripped Out The downloadable checkpoint is text-only, with forced thinking. The hosted qwen3.8-max API keeps the multimodal /glossary/multimodal input and the 1M- token /glossary/token context that made the product notable at launch. Same name, different artifact. If you wanted the model that reviews images and reads million-token documents, you still have to pay per token. That gap is the first signal about what's really being given away. Alibaba shipped a competitive text model as a marketing artifact while keeping the differentiated capabilities locked to the API. It's a preview, not the product. The License Is the Real Story The bigger issue is the license. It's a custom Qwen3.8-Max License - not Apache 2.0, not MIT. The terms that matter: First, you must keep the copyright and permission notice in all copies. That's routine. Second, if your product exceeds 100 million monthly active users or 20 million dollars in monthly revenue, you must display the model name prominently in your UI. Third - and this is the one that changes things - if you run a model-as-a-service or AI work assistant business with over 50 million dollars in trailing-twelve-month aggregate revenue including affiliates, you need a separate paid license before commercial use. Purely internal use is carved out, provided you don't expose the model or its capabilities to third parties. For most teams, none of those thresholds bite. For anyone building a product on top of it and selling access, they do. That's the exact distinction Apache 2.0 does not make and this license does. The Sibling That Got the Clean License Compare it to the sibling release. Qwen3.8-27B landed August 14 under plain Apache 2.0 - no thresholds, no revenue riders - and it scored 61.7 on SWE-Bench Pro, ahead of Claude /glossary/claude Opus 4.6 Max. The small model got the clean license. The flagship got the revenue-share. That split is a strategy, not an accident. Alibaba is giving away the model that builds developer goodwill and charging for the one that builds businesses. Open weights become a funnel: try the small model free, then license the big one when you start making real money. What Open Means Now The Qwen 3.8-Max license is the latest move in a redefinition that's been running all year. Meta ships genuinely open models on Apache. Alibaba ships tiered openness with revenue triggers. The word open now has a spectrum, and the license text matters more than the marketing. For developers, the practical lesson is simple: read the license before you build, not after. A model that's open today under a free license can be the same model that demands a revenue share tomorrow, once your product crosses a threshold you never noticed on launch day. Sources: Qwen3.8-Max Hugging Face repository and license, August 12, 2026; AI Tools Recap daily briefing, August 17, 2026; Nvidia deployment engineering blog, August 12, 2026. Get AI news in your inbox Daily digest of what matters in AI. Key Terms Explained Claude /glossary/claude Anthropic's family of AI assistants, including Claude Haiku, Sonnet, and Opus. Hugging Face /glossary/hugging-face The leading platform for sharing and collaborating on AI models, datasets, and applications. Multimodal /glossary/multimodal AI models that can understand and generate multiple types of data — text, images, audio, video. NVIDIA /glossary/nvidia The dominant provider of AI hardware.