On-Prem AIFree referencePDF, Excel, Word export

Open-Weight LLM License Terms Reference

Before an open-weight model touches your ERP data, someone has to confirm you are allowed to run it for commercial work. This reference lists the license typically attached to each popular family and the conditions people most often miss.

Download:

Model family	Publisher	License (typical)	Commercial use	Conditions to check	On-prem ERP note
Llama 3.1 / 3.3	Meta	Llama Community License	Yes, with conditions	Attribution (Built with Llama), acceptable use policy, separate license needed above 700M monthly active users, derivative model naming rule	Fine for internal ERP assistants; keep the AUP and attribution text in your documentation
Llama 4	Meta	Llama 4 Community License	Yes, with conditions	Similar structure to earlier Llama licenses; read the current text for any regional restrictions on multimodal use	Have counsel read the exact license for your country before deploying
Qwen3 (most sizes)	Alibaba Qwen	Apache 2.0	Yes	Keep license and notices; verify the model card for each size	Permissive; common choice for air-gapped deployments
Qwen2.5 (mixed)	Alibaba Qwen	Apache 2.0 for most sizes; some sizes use the Qwen license or a research license	Depends on size	Check the LICENSE file of the exact checkpoint, since sizes differ	Record the exact checkpoint and its license in your model inventory
Mistral 7B	Mistral AI	Apache 2.0	Yes	Keep license and notices	Small and cheap to host; limited on long ERP documents
Mixtral 8x7B	Mistral AI	Apache 2.0	Yes	Keep license and notices	Mixture of experts, needs memory for all experts
Mistral Small (recent open releases)	Mistral AI	Apache 2.0	Yes	Verify per release; some Mistral models use other licenses	Good size for a single-GPU private assistant
Mistral Large 2	Mistral AI	Mistral Research License	No, research use only without a commercial agreement	Commercial deployment needs a separate agreement with the publisher	Do not deploy in production without that agreement
Codestral 22B	Mistral AI	Mistral AI Non-Production License	No for production	Non-production and research use only unless licensed separately	Check before using for ABAP or SQL code generation inside a company
Gemma 2 / Gemma 3	Google	Gemma Terms of Use	Yes, with conditions	Prohibited use policy applies and must be passed to downstream users; not an OSI open source license	Fine for internal use; review redistribution terms if you ship it to customers
gpt-oss-20b / gpt-oss-120b	OpenAI	Apache 2.0 with a usage policy	Yes	Keep license; the accompanying usage policy applies	Open-weight reasoning models that run on-prem; check VRAM needs first
DeepSeek-R1	DeepSeek	MIT	Yes	Keep license notice. Distilled variants inherit the license of the base model (Qwen or Llama)	Check the base model license for any distilled checkpoint you use
DeepSeek-V3	DeepSeek	Code under MIT; weights under the DeepSeek Model License	Yes, with use restrictions	Model license lists prohibited uses	Very large; needs a multi-GPU node
Phi-3 / Phi-4	Microsoft	MIT	Yes	Keep license notice	Small models suited to edge and plant-floor devices
Granite	IBM	Apache 2.0	Yes	Keep license and notices	Enterprise-oriented small and mid-size models
OLMo	Allen Institute for AI	Apache 2.0	Yes	Keep license and notices	Fully open training data and code, useful where auditability matters
Command R / R+ (open weights)	Cohere	CC BY-NC 4.0 with acceptable use addendum	No for non-commercial license	Commercial use requires a commercial agreement	Not usable for production without that agreement
Nemotron	NVIDIA	NVIDIA Open Model License	Yes, with conditions	Read the current license for guardrail and redistribution terms	Pairs naturally with NVIDIA inference stacks

Licenses change between releases and between sizes of the same family. Always read the LICENSE file shipped with the exact checkpoint you download. This table is a starting point for your review, not legal advice.

Model and license review

An optional walkthrough of which open-weight model fits your hardware, data and license constraints.

Optional. The tool and every export work without this. No spam.

How to use this table

Search by model family or by license name. Read the Commercial use column first to rule models out, then the conditions column for attribution, acceptable use policies and user-count thresholds. Add the exact checkpoint name and license to your internal model inventory so that an auditor or a customer questionnaire can be answered in minutes. Export the filtered rows to Excel or PDF to attach to a procurement or architecture decision record.

Open weights are not the same as open source

A model with downloadable weights can still carry restrictions. Apache 2.0 and MIT are permissive. The Llama and Gemma licenses are custom terms with acceptable use policies. Research or non-production licenses forbid use in a live business process. The distinction matters because an internal ERP assistant that answers real purchasing or production questions is a commercial use even if no customer ever sees it.

Permissive: Apache 2.0, MIT
Custom community licenses: attribution and policy flow-down
Non-commercial or research only: needs a separate agreement
Distilled models inherit the base model license

Common mistakes

Teams often judge a family by one famous release and assume every size shares its license. Others forget that a fine-tuned derivative keeps the obligations of its parent. A third mistake is ignoring the acceptable use policy, which can prohibit use cases that touch regulated areas. Finally, license text can be updated in later releases, so a version you approved last year may not match the one you download today.

Frequently Asked Questions

Can I use Llama commercially for an internal ERP assistant?

Generally yes under the Llama Community License, subject to its conditions such as the acceptable use policy and attribution. The 700 million monthly active user threshold rarely matters for a company-internal tool. Read the exact text for the release you use.

Does Apache 2.0 mean there are no restrictions at all?

No. Apache 2.0 allows commercial use, modification and redistribution, but you must keep the license and notice files and state significant changes. It also includes a patent grant. Some models add a separate usage policy on top of it.

Are distilled DeepSeek models covered by the MIT license?

The DeepSeek-R1 release is MIT, but distilled checkpoints are built on Qwen or Llama bases and also inherit those base licenses. Check both before approving a specific checkpoint for production.

How often should we recheck licenses?

At each model upgrade and at least annually. Publishers sometimes change terms between versions. Store the license file and download date alongside the model weights in your artifact repository.

Netray deploys open-weight models on customer GPUs and can help you document the license, hardware and data flow for each one. We do not give legal advice, so counsel should confirm the final reading.

Talk to an Expert Explore More Free Tools

Related Tools

Developer Tools

More On-Prem AI tools

See all

Back to All Tools