StableLM’s Context Window and Stability AI’s Licensing Terms

StableLM’s context window is short by current standards and has been throughout: 4,096 tokens across the releases that matter. The licence is the part that has moved, and it has moved into a shape — free below a revenue threshold, paid above it — that requires you to know your own company’s numbers before you can answer whether you may use the model.

The StableLM releases

• StableLM-Alpha (April 2023). The first release, at 3B and 7B, announced as an alpha with larger sizes promised. A v2 of the Alpha models followed with a retrained recipe.
• StableLM-3B-4E1T (September 2023). A 3B model trained for four epochs over one trillion tokens, which is what the name encodes. This was the technically most substantial of the early releases and came with a technical report.
• StableLM Zephyr 3B. A preference-tuned chat variant built on the 3B base.
• StableLM 2 (2024). A retrained generation at 1.6B and 12B, multilingual, with base and instruct variants.
• Stable Code and Stable LM Code variants. Code-focused models with their own cards, released alongside.

Stability AI’s language-model work has always sat alongside a much larger image-generation programme, and the licensing decisions described below were made for the company as a whole rather than for the language models specifically. That is worth knowing because the reasoning in the public discussion of those terms is usually about Stable Diffusion.

Context length by release

StableLM-Alpha (3B / 7B)        4,096 tokens
StableLM-Alpha v2               4,096 tokens
StableLM-3B-4E1T                4,096 tokens
StableLM Zephyr 3B              4,096 tokens
StableLM 2 (1.6B / 12B)         4,096 tokens

A consistent 4,096 across the line is unusual enough to be worth a sentence. It reflects when these models were designed — a 4K window was the standard of the 2023 generation — and the fact that Stability did not ship a long-context variant of the kind 01.AI released for Yi. If you need a long window from a small open model, StableLM is not the line to look at.

The one place StableLM 2 is genuinely interesting on this axis is throughput per unit of context rather than the context itself. A 1.6B model with a 4K window keeps a very small KV cache — the cache scales with layer count, head count and sequence length, and all three are small here — so you can run a high concurrency on modest hardware. That is the shape of workload these models were built for: many short requests on a device or a single GPU, not one long document. Choosing StableLM and then wishing it had a longer window is choosing against the thing it was optimised for.

Three licensing regimes

Creative Commons, 2023

The original StableLM-Alpha weights were released under CC BY-SA-4.0 — a Creative Commons licence, not a software licence. StableLM-3B-4E1T used CC BY-SA-4.0 as well. This was an unusual choice and it has a consequence people frequently miss: SA means share-alike. A share-alike licence requires derivative works to be released under the same terms. Whether a fine-tune of a CC BY-SA model is a derivative work is an unsettled question, and the uncertainty itself is the problem for anyone planning to build on it commercially. Some later variants were released under CC BY-NC-SA, adding a non-commercial term and removing the ambiguity in the least convenient direction.

Non-commercial research licence

StableLM 2 was initially released under a Stability AI Non-Commercial Research Community License, with commercial use requiring a separate agreement with Stability. This applied to several of the 2024 models across Stability’s catalogue.

The Community License, from July 2024

Stability AI then introduced a Community License covering its models generally. Its structure: free for non-commercial use, and free for commercial use by individuals and organisations with annual revenue below a stated threshold of one million US dollars. Above that threshold, commercial use requires an enterprise agreement. Stability publishes the current text at stability.ai/license, and that text governs rather than any summary of it, including this one.

The direction of that sequence is the thing to notice. Stability went from a Creative Commons licence, through an outright non-commercial restriction, to a threshold model — three different answers to “may I use this commercially” in roughly fifteen months, all applied to weights people had already downloaded. That is not a criticism of any one choice; it is an argument for recording, at the moment you take a model, which terms were attached to it. A licence published at a URL can be revised. A file you saved cannot.

The revenue threshold, and what it applies to

A revenue-conditioned licence is a different animal from the conditional-but-flat licences used by Falcon and DBRX, and from the unconditional Apache 2.0 that Mistral and Qwen use for their open releases. Three practical points follow.

• The condition is on you, not on the use. The same deployment is permitted for one company and not for another, based on a number that has nothing to do with the model. That means your licence status can change without you doing anything — a good year crosses the threshold.
• It is total revenue, not revenue from the model. Read the definition carefully; a threshold defined on organisation-wide annual revenue catches a large company making trivial use of the model.
• It is not an open-source licence. A field-of-use or entity restriction fails the Open Source Definition, so “open” here means downloadable weights. If your dependency review distinguishes these, StableLM under the Community License belongs in the same bucket as Meta’s Llama terms — see Llama’s commercial-use conditions, which use a monthly-active-user threshold instead of a revenue one but are structurally the same idea.

A last point about StableLM specifically, which is really a point about any model from a company under commercial pressure. Stability’s language models were released alongside a much larger and much more contested image programme, and the licensing changes tracked the company’s commercial situation rather than anything about these weights. If you are building on a small model from a venture-funded lab, the relevant risk is not that the model degrades — a downloaded checkpoint is permanent — but that the terms under which you may use it, and the support around it, are attached to an entity whose plans can change. Downloading and archiving the weights and the licence text together is the whole mitigation, and it costs nothing.

What to verify

1. Identify the exact repository and its release date. StableLM spans three licensing regimes and the date is the strongest signal for which one applies.
2. Read the LICENSE or LICENSE.md file in that repository. If it references a URL rather than embedding the text, note that the terms at that URL can change after you download.
3. If the Community License applies, get the revenue threshold and its definition read by whoever owns that question in your organisation, and record the date you read it.
4. For CC BY-SA models, get a view on whether your fine-tune is a derivative work before you distribute it, not after.
5. Pin the commit hash. It is the only durable evidence of which terms were attached to the weights you actually took.