Skip to content

Model licenses

The model weights served by ai-session are governed by the license each model publisher attached to them. When you run a private session for yourself, you are the licensee using the model. When you share your session's access key with labmates so they reach it over their own tunnels — the sharing model described on Coding Sessions and in the command reference — you are making the model available to other people, which is where the licenses differ. This page states, per model, what the license is, where its authoritative text sits on disk, and what you must do when you serve it to others. It is a practical summary, not legal advice; the on-disk license file is the controlling text in every case.

License at a glance

In the table, <models> is the service's model store, $AISESSION_HOME/models after module load ai-session.

Model key License On-disk license file What serving to others requires
qwen2.5_coder_32B Apache-2.0 <models>/Qwen2.5-Coder-32B-Instruct/LICENSE Nothing beyond keeping the license and any NOTICE with redistributed weights.
qwen3_4b Apache-2.0 <models>/Qwen3-4B/LICENSE Same as above.
qwen3_32B Apache-2.0 <models>/Qwen3-32B/LICENSE Same as above.
qwen2.5_0.5B Apache-2.0 <models>/Qwen2.5-0.5B-Instruct/LICENSE Same as above.
qwen2.5_72B Qwen (Tongyi) community license <models>/Qwen2.5-72B-Instruct/LICENSE Retain the Qwen attribution notice; observe the "Built with Qwen" and large-scale-use terms below.
llama3.1_70B Llama 3.1 Community License + Acceptable Use Policy <models>/Meta-Llama-3.1-70B-Instruct/LICENSE and .../USE_POLICY.md Provide the license, display "Built with Llama", follow the Acceptable Use Policy, and record acceptance via the acknowledgment gate.

Models on the roadmap (see Command Reference) will be added to this table with their licenses when they become servable: qwen3.5_122B (Qwen3.5-122B-A10B) is Apache-2.0; glm5.2_753B (GLM-5.2) is MIT (<models>/GLM-5.2-FP8/LICENSE); the GLM-5.1 license terms will be recorded here before that model is offered.

The Apache-2.0 models

qwen2.5_coder_32B, qwen3_4b, qwen3_32B, and qwen2.5_0.5B are released under the Apache License, Version 2.0. Apache-2.0 is a permissive open-source license: it places no restriction on serving the model to other users and imposes no in-product attribution requirement. The only standing obligation is that if you redistribute the weights themselves (copy them elsewhere, not merely serve inference from them), you keep the LICENSE file and any accompanying NOTICE with the copy. Serving these models through ai-session, whether to yourself or to your lab, needs nothing further.

The Qwen 72B community license

qwen2.5_72B (Qwen2.5-72B-Instruct) is released under the Qwen LICENSE AGREEMENT, a community license from Alibaba Cloud rather than an OSI-approved open-source license. It permits research and commercial use but attaches obligations that matter once the model is offered to others as a service:

  • Attribution. Copies you distribute must retain, in a Notice text file, the exact line the license specifies: "Qwen is licensed under the Qwen LICENSE AGREEMENT, Copyright (c) Alibaba Cloud. All Rights Reserved."
  • "Built with Qwen". If you use the model's outputs to train, fine-tune, or otherwise improve another AI model that you then distribute or make available, you must prominently display "Built with Qwen" or "Improved using Qwen" in that product's documentation.
  • Changed files. If you modify the materials, the modified files must carry prominent notices stating that you changed them.
  • Large-scale commercial use. If you use the materials commercially in a product or service with more than 100 million monthly active users, you must request a separate license from Alibaba Cloud. This threshold is far above any RCC lab setting and is noted only for completeness.
  • Trademarks. No trademark license is granted beyond what is needed to satisfy the attribution requirement above.

When you host qwen2.5_72B for your lab through ai-session, the practical duty is the attribution notice; the full and controlling terms are in the on-disk LICENSE above.

The Llama 3.1 license

llama3.1_70B (Meta-Llama-3.1-70B-Instruct) is released under the Llama 3.1 Community License Agreement, accompanied by a separate Acceptable Use Policy (USE_POLICY.md). Like the Qwen license, it is a community license with conditions, and it is the most restrictive of the models staged here. Serving it to others carries these obligations:

  • Provide the license. If you distribute or make the Llama materials (or a product using them) available to a third party, you must provide a copy of the Llama 3.1 Community License Agreement with it.
  • "Built with Llama". You must prominently display "Built with Llama" on a related website, interface, or documentation when you make the model available.
  • Naming. If you use Llama to create, train, or improve another AI model that you distribute, its name must begin with "Llama".
  • Attribution. Retain, in a Notice file, the line the license specifies: "Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved."
  • Acceptable Use Policy. Your use, and the use by anyone you serve, must comply with the Acceptable Use Policy in USE_POLICY.md (which prohibits a specific list of harmful uses).
  • Large-scale commercial use. If the product or service using Llama exceeded 700 million monthly active users on the release date, you must request a license from Meta. As with Qwen, this is noted only for completeness.

Because these duties bind whoever offers the model, ai-session refuses to serve llama3.1_70B until you have acknowledged them, as described next.

Serving Llama 3.1: the acknowledgment gate

llama3.1_70B is served to any user — the Community License permits this use, subject to the conditions above — but the first attempt to serve it is refused until you record that you accept the license. This is a deliberate, non-interactive gate so that scripts can satisfy it once and proceed: the serving command is refused until it is run with ACCEPT_LLAMA_LICENSE=1 set in the environment. The first time:

ACCEPT_LLAMA_LICENSE=1 ai-session chat --model llama3.1_70B

Setting ACCEPT_LLAMA_LICENSE=1 writes a one-time acceptance record to your per-user state directory at <state-dir>/logs/licenses/<user>_llama3.1_70B.accepted. The record holds the timestamp, your username, the model key, and the on-disk license path you accepted. Once it exists, later starts reuse it and need no environment variable.

Without the acknowledgment, and with no record already on file, the start is refused before any GPU is reserved, and the refusal prints the on-disk license paths and the variable to set.

The Apache-2.0 models are permissive and are not gated; qwen2.5_0.5B is a smoke-test checkpoint (not a user model) and needs no license acknowledgment. qwen2.5_72B is served and is not license-gated, but its attribution obligation above still applies when you host it for others.

Reading the authoritative text

Every license summarized here is quoted in full in the on-disk file listed in the table. To read one directly on a login node (after module load ai-session), for example the Llama license and its Acceptable Use Policy:

less "$AISESSION_HOME/models/Meta-Llama-3.1-70B-Instruct/LICENSE"
less "$AISESSION_HOME/models/Meta-Llama-3.1-70B-Instruct/USE_POLICY.md"

If in doubt about an obligation for your specific use, read the on-disk license, which controls, and raise questions about service policy with RCC staff.