Qwen3.8-2.4T-A95B: above fifty million dollars a separate licence is required, not for internal use
7 min read
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 — 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. 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. 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 are not sovereignty. They are a promise written into a licence file.
See the service · Talk to an engineer
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, unlike cases where a permissive licence sits next to a usage policy that hollows it out. 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.
How we solve this
Comply: the register of models in production — licence archived in the wording of the day it was downloaded, which conditions apply and at what threshold, which entities in the group count toward the affiliates’ calculation, internal use or exposed to third parties — with the dated trail ready to show an inspector, a client in a tender, or a board.
Decide: the same system holds models, data, contracts, archives and documents together in a single operating model, on which AI agents execute decisions with a human operator in command, for large enterprises, defence, government and healthcare. Internal use inside the client’s own perimeter is exactly the case this licence exempts: it is how we build, without reselling access to the model to third parties as a service. Always in two delivery modes — on-premise on self-contained machines with no deep integration into the client’s network, or dedicated cloud with a dedicated VPN and a data centre in Italy — and always with shared governance.
From the first session, at no cost, comes the dated register of the models you have in production: licence archived, applicable conditions and thresholds, internal use or exposure to third parties — including the boxes left blank. It stays with you even if we do not carry on together.