The inference runtime
Licensing
Model weight licences, the lite revenue cap, and the Apache-2.0 swap-in
The code in this repository is MIT (LICENSE).
The model weights are not. They are third-party works, downloaded at runtime rather than shipped inside the image, and each carries its own licence. Those terms attach to you, the operator.
Qwen3-VL-8B-Instruct: quality profile
Apache-2.0. No revenue cap, no field-of-use restriction, commercial use fine. Source: Qwen/Qwen3-VL-8B-Instruct-GGUF. This is the recommended path for commercial use.
Qwen3-VL-2B-Instruct: lite-apache profile
Apache-2.0. Source: Qwen/Qwen3-VL-2B-Instruct-GGUF.
LFM2.5-VL-1.6B: lite profile, revenue-capped
lite uses LiquidAI LFM2.5-VL-1.6B under the LFM Open License v1.0 (licence text, in the model repo). It is Apache-2.0-shaped with one added condition, verbatim:
5. Commercial Use Limitation. (a) The rights granted under this License for Commercial Use are conditioned upon You or Your Legal Entity not exceeding the Threshold. (b) Any Commercial Use of the Work or a Derivative Work by a Legal Entity that exceeds the Threshold is not licensed under this Agreement. (c) The Threshold shall not apply to a Qualified Non-Profit Organization's use of the Work or a Derivative Work for Non-Commercial or Research Purposes.
and the definition it turns on:
"Threshold" shall mean annual revenue of 10 million United States dollars ($10,000,000) or more.
with "Commercial Use" defined as "any use of the Work for direct or indirect commercial advantage or monetary compensation", and "Legal Entity" including parents, subsidiaries and anything under common control.
The $10M revenue cap binds the operator, not this project:
- Running
liteat home, for a household, or for research: no cap issue, that is not Commercial Use. - A company under $10M annual revenue using it commercially: permitted.
- A company at or over $10M annual revenue:
liteis not licensed to you. Nothing in this repository changes that; the condition attaches to your use of the weights. - A 501(c)(3) (or foreign equivalent) using it non-commercially: exempt from the Threshold by clause 5(c).
This is a licence summary, not legal advice. If the cap might bind you, read the licence and ask your own lawyer.
The Apache-2.0 swap-in
Every model slot has an Apache-2.0 alternative:
| slot | revenue-capped default | Apache-2.0 alternative |
|---|---|---|
quality (GPU flagship) | none | already Apache-2.0 (Qwen3-VL-8B-Instruct) |
lite (small/CPU) | LFM2.5-VL-1.6B | Qwen3-VL-2B-Instruct → MODEL_PROFILE=lite-apache |
-e MODEL_PROFILE=lite-apache # the entire migrationNote on that swap: we measured LFM2.5-VL-1.6B and we did not measure Qwen3-VL-2B. It is the official Apache-2.0 Qwen3-VL sibling in the same size class, from the same family as the quality model, with first-class llama.cpp support and an official GGUF release. We expect it to be at least comparable and we have not proven it. If accuracy matters more than the small-model convenience, quality is the measured Apache-2.0 answer.
Other components
llama.cpp (MIT): used as the official upstream container image, not vendored.
Food data licences: USDA FoodData Central is public domain; Open Food Facts is ODbL and share-alike; the lowcarbcheck API is remote-only because BLS 4.0 forbids redistribution. See Food data.