--- title: "Qwen3.8-2.4T-A95B: above fifty million dollars a separate licence is required, not for internal use" description: "The Qwen3.8-2.4T-A95B model on Hugging Face: the weights’ licence requires a separate licence above $50 million in revenue, but exempts internal use." url: https://csidia.com/en/notes/qwen-2-4t-licenza-soglia-cinquanta-milioni-uso-interno/ lang: en date: 2026-08-16 label: "Observatory" publisher: CSIDIA S.r.l. (https://csidia.com/) italiano: https://csidia.com/note-operative/qwen-2-4t-licenza-soglia-cinquanta-milioni-uso-interno/ --- # Qwen3.8-2.4T-A95B: above fifty million dollars a separate licence is required, not for internal use **`Qwen/Qwen3.8-2.4T-A95B`** appeared on Hugging Face on 8 August 2026: the API’s `createdAt` field reads **2026-08-08T01:50:52Z**. We checked it on 16 August, redownloading `LICENSE` and `README.md`. At that moment: **7,932 downloads**, **986 likes**, **224 files**, of which **213 `.safetensors` shards**. The tags include `qwen3_5_moe_text`, `text-generation`, `license:other`, `region:us`; the README’s front matter declares `license: other`, `license_name: qwen3.8-max`, `license_link: LICENSE`. The name already states the architecture: **2.4 trillion parameters in total, 95 billion active** — a mixture of experts, 512 experts, 10 routed plus 1 shared per token. ## The licence, read in full It is not Apache 2.0. It is a licence of its own, the “Qwen3.8-Max License”, “Copyright (c) 2026 Qwen”. The grant names the weights explicitly: “Permission is hereby granted, free of charge, to any person obtaining a copy of this software, including the model weights, parameters, configuration files, inference code and associated documentation files (collectively, the ‘Software’), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, sell, deploy, host, fine-tune, and create derivative works from […] copies of the Software; and to permit persons to whom the Software is furnished to do so, subject to the following conditions”. Two conditions, not one. The first concerns the visibility of the name above a threshold of users or revenue: “If the Software […] is Used for any of the licensee’s commercial products or services that have more than 100,000,000 monthly active users or US$ 20,000,000 (or equivalent in other currencies) monthly revenue, respective model name must be prominently displayed on the user interface of such product or service”. The second is the one that matters. “If the licensee or any of its affiliates conducts a Model as a Service or AI Work Assistant business, and the aggregate revenue of the licensee and its affiliates exceeds US$50,000,000 (or the equivalent amount in any other currencies) during any consecutive twelve (12) months, the licensee shall obtain a separate license from Qwen before Using the Software or its derivative works for any commercial purpose. The foregoing requirement shall not apply to the licensee’s internal Use of the Software, provided that such Use does not make the Software, its outputs, or its underlying model capabilities available to any third party.” In plain terms: once the licensee’s group crosses fifty million dollars of aggregate revenue in any rolling twelve months while running a Model-as-a-Service or AI Work Assistant business, a separate licence from Qwen becomes a precondition for further commercial use — unless the use stays internal and nothing built on the model is handed to a third party. Two definitions decide when it applies. “‘Model as a Service’ means giving a third party access to language model inference or fine-tuning (e.g., via API or a hosted endpoint) in a manner that allows such third parties to exercise meaningful control over the inputs, parameters, or training data. This does not include the mere relaying of requests to models hosted by other third parties.” And: “‘AI Work Assistant’ means an independent AI-powered product primarily designed for AI-assisted coding or office productivity (e.g., Qoder and QwenWork). It does not include: (a) a single-purpose AI tool […]; (b) an AI assistant primarily designed for a domain other than coding or office productivity […]; or (c) an AI assistant that is a feature of a product whose primary purpose is not AI-assisted coding or office productivity.” It closes with the “AS IS” clause and: “THE USE OF THE SOFTWARE MUST COMPLY WITH APPLICABLE LAWS AND REGULATIONS, AND MUST NOT INFRINGE THE INTELLECTUAL PROPERTY RIGHTS OF ANY THIRD PARTY.” It also gives a contact for anyone who needs to negotiate: model-business@notice.qwencloud.com. ## A departure from Qwen’s usual practice The most recent Qwen releases we checked on the Hugging Face API — `Qwen3-Next-80B-A3B-Instruct`, `Qwen3-Coder-480B-A35B-Instruct`, `Qwen3-235B-A22B`, `QwQ-32B` — all declare `license: apache-2.0`. Qwen3.8-2.4T-A95B is the exception: it is the flagship model, the only one carrying revenue thresholds. The same mechanism, in different form, [appears in LFM2.5](https://csidia.com/en/notes/lfm-open-license-soglia-dieci-milioni) — there, ten million dollars of the licensee’s annual revenue; here, fifty million across the group for anyone running Model as a Service. A broad permission, conditioned on a number that changes from one text to the next. ## The open weights are a reduced version The README says so without ambiguity: “In particular, **Qwen3.8-Max** is the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools”. Qwen3.8-Max, the version served via API, adds vision input, a non-thinking mode, a one-million-token context by default and built-in tools, compared with the weights downloaded here, and points anyone seeking managed inference to “the official Qwen API service is provided by Qwen Cloud.” This is not a hidden shortcoming: it is stated plainly, [a declared commercial choice, much like others in the sector](https://csidia.com/en/notes/licenza-permissiva-usage-policy-che-vieta). Anyone comparing a benchmark of the open weights with the hosted experience under the same name is comparing two different products. ## The question almost no one asks 213 shards of weights, for 2.4 trillion total parameters and 95 billion active. The adoption numbers tell of a gap: 986 likes against 7,932 downloads — a model watched and discussed, but downloaded in modest proportion next to other, more manageable open models we have covered in these same weeks, [as with the opposite case of a small model with no declared licence at all](https://csidia.com/en/notes/qwen3-8-huggingface-download-senza-licenza). We do not know, and we do not claim, what infrastructure is needed to run a model of this size: it is not a figure a Hugging Face card provides. But the question deserves asking, because it is the one the licence chapter risks pushing out of view: [weights that can be downloaded without the sizing work that makes them runnable inside your own perimeter](https://csidia.com/en/hardware-and-open-weight-models) are not sovereignty. They are a promise written into a licence file. [See the service](https://csidia.com/en/solutions/ai-governance) · [Talk to an engineer](https://csidia.com/en/contact) ## What we don’t know We have not verified what hardware is needed to run the model, and we do not claim to. We offer no legal interpretation of the thresholds, nor do we determine whether an activity amounts to “Model as a Service”: we report the definitions verbatim; applying them to a real case is a matter for a lawyer. We have not assessed the model’s quality and do not report the benchmarks on the card. Among the 224 files, beyond the licence and the README, there are only technical files — tokenizer, chat template, shard index: [no second restrictive document sits alongside the weights](https://csidia.com/en/notes/deepseek-v4-pro-mit-senza-usage-policy), unlike [cases where a permissive licence sits next to a usage policy that hollows it out](https://csidia.com/en/notes/minimax-h3-licenza-esclude-unione-europea). Qwen publishes its own terms for Qwen Cloud, updated as of 1 April 2026: they govern the managed service, not the repository. A licence is verified on the date it is downloaded: it should be archived alongside the weights. ## Sources - [Hugging Face — Qwen/Qwen3.8-2.4T-A95B, model card](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B) - [Hugging Face — full text of the LICENSE file](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/LICENSE) - [Hugging Face — model API (metadata, files, tags)](https://huggingface.co/api/models/Qwen/Qwen3.8-2.4T-A95B) Note by CSIDIA S.r.l. (Vignola, MO, Italy): https://csidia.com/en/notes/qwen-2-4t-licenza-soglia-cinquanta-milioni-uso-interno/ · Who we are and what we do: https://csidia.com/en/at-a-glance/ · engagement@csidia.com