Qwen3.8-2.4T-A95B: the weights are open, the licence is not

Qwen
Coding Agents
Verification

Qwen has published the weights for Qwen3.8-2.4T-A95B on Hugging Face, per the model card the first open release of a Max-class model. The files really are there: 213 safetensors files totalling roughly 4.9 terabytes, plus config, index, tokenizer, and chat template. The notable part is still not the size. It is the file next to them. It is called LICENSE, it is not Apache or MIT, and it contains revenue thresholds.

What is actually in the repository?

A complete model, not a placeholder. Qwen3.8-2.4T-A95B is a mixture-of-experts model with 2.4 trillion parameters, of which 95 billion are active per token. The model card names 92 layers, a hidden dimension of 8192, and 512 experts, 10 routed plus one shared per token. The layout alternates between Gated DeltaNet and Gated Attention: 23 repetitions of three DeltaNet blocks followed by one attention block, each with MoE. Context length is 262,144 tokens natively and, according to Qwen, extensible to 1,010,000 tokens. The card is explicit about two limits: the open model is text only, and thinking cannot be disabled. Every response begins with a <think> block. Reasoning depth is controlled through reasoning_effort with the levels xhigh, medium, and low.

What does the licence say?

Not a standard licence, but the “Qwen3.8-Max License”. It grants a great deal up front: use, copy, modify, publish, distribute, sublicense, sell, deploy, host, fine-tune, and create derivative works. Two conditions are attached, and both only bite above a certain size.

Attribution in the interface. If the software is used in a commercial product or service with more than 100 million monthly active users or more than US$20 million in monthly revenue, the model name must be prominently displayed in that product’s user interface.
Separate licence required. If the licensee or any affiliate runs a Model as a Service or AI Work Assistant business and their aggregate revenue exceeds US$50 million over any consecutive twelve months, a separate licence from Qwen is required before any commercial use. Internal use is exempt, provided the software, its outputs, and its underlying capabilities are not made available to any third party.

Both terms are defined in the licence. “Model as a Service” means giving a third party access to inference or fine-tuning, for instance via an API or a hosted endpoint, in a way that lets that third party exercise meaningful control over inputs, parameters, or training data. Merely relaying requests to models hosted by others does not count. “AI Work Assistant” means an independent product built primarily for AI-assisted coding or office productivity, with Qoder and QwenWork named as Qwen’s own examples. Single-purpose tools, assistants in other domains, and assistant features inside products with a different primary purpose are explicitly excluded.

The Hugging Face metadata field says only license: other with the name qwen3.8-max. The tag tells you nothing. The conditions live in the file, and anyone building open weights into a product has to read that file.

Is the open model the same as Qwen3.8-Max?

No. The announcement carries the name Qwen3.8-Max, the release is Qwen3.8-2.4T-A95B. Qwen itself describes Qwen3.8-Max as the official version based on it, with additional features: vision input, a non-thinking mode, one million tokens of context by default, and built-in tools. Run the weights yourself and you get the foundation, not the service. That is not a complaint, just an expectation worth correcting before you plan around it.

What do the benchmarks say?

They are vendor figures from the model card, and the measured column is called Qwen3.8-Max, the cloud version, not the open artefact. The comparison runs against Opus 4.8, Fable 5, GPT 5.6 Sol (max), and the predecessor Qwen3.7-Max. On Terminal Bench 2.1 the numbers are 86.6 against 84.6 for Opus 4.8, 84.6 for Fable 5, and 88.8 for GPT 5.6 Sol. On SWE-bench Pro it is 67.7 against 69.2, 80.0, and 64.6. On DeepSWE 1.1, Qwen3.8-Max sits behind all three at 56.6. GPQA Diamond: 92.6. The strongest entry in the table is PaperBench at 93.0.

The footnotes matter more than the numbers. Terminal Bench 2.1 was measured for Qwen with Claude Code as the harness, avg@10, with a five-hour timeout; for the comparison models, the best published score across harnesses is quoted. On SWE-bench Pro, Qwen states that problematic tasks were corrected and all baselines re-evaluated on the refined benchmark. Several rows are in-house benchmarks: QwenSWEBench, QwenQoderBench, QwenReactBench, QwenSVGBench, CoWorkBench. On SkillsBench the external models run under Claude Code or Codex while the Qwen models run under OpenCode. All of it is documented openly, and all of it is still the vendor measuring itself.

What does this mean for your team?

Self-hosting stays unrealistic for most. 4.9 terabytes of weights is not a server, it is a cluster; Qwen recommends SGLang, vLLM, or TokenSpeed for production load and otherwise points at its own cloud API. The value of open weights sits elsewhere, and with Kimi K3 the observation was the same: third parties can re-measure what the vendor claims. This time a second check comes with it, and that one is legal. If your product passes inference through to third parties, or is a coding or office assistant, the US$50 million threshold belongs in the decision before the model moves into your architecture. For an internal tool whose outputs never leave the building, the licence text puts that threshold out of scope.

The division of labour does not change. A model with 262,144 tokens of context earns its place on long, tedious stretches: large repositories, terminal work, research across many documents. The sign-off stays with you, and as described in agentic engineering, reliability comes from the architecture around the model, not from a row in a table. Open weights lower one barrier. This licence adds a new one, and it is not in the metadata field.