License
tensorless is open source under the Apache License, Version 2.0. This page says what that covers, what keeps another license, and where each notice is.
What is Apache-2.0
Section titled “What is Apache-2.0”The launcher, the recipes, the container files, these docs and every change made for tensorless are licensed under Apache-2.0. The project’s NOTICE file carries the attribution that Apache-2.0 asks you to keep.
What keeps its own license
Section titled “What keeps its own license”The engine’s upstreams. The engine in engine/ is derived from MIT-licensed projects, and their copyright and permission notices stay with it:
| Project | License |
|---|---|
| llama.cpp / ggml, the ggml authors | MIT |
| llama-cpp-turboquant, TheTom | MIT, the same notice |
| llamAmpere, JakeATX | MIT, the same notice |
| exllamav3, turboderp (portions, for the EXL3 weight types) | MIT |
Libraries vendored into the engine, such as cpp-httplib, nlohmann/json and xxHash, each under its own license (MIT, BSD-2-Clause, public domain and others). The launcher’s one dependency, BurntSushi/toml, is MIT. The CUDA libraries in the engine images come from NVIDIA’s CUDA base images, under the NVIDIA CUDA Toolkit license.
THIRD_PARTY_NOTICES.md lists every one of them and where its notice is. The license texts are in LICENSES/ and in the engine’s own license files. The release archives of the launcher carry LICENSE, NOTICE, THIRD_PARTY_NOTICES.md and LICENSES/, and each engine image carries the same files under /opt/tensorless/.
Model weights
Section titled “Model weights”Model weights are not part of tensorless. They are downloaded from Hugging Face and keep their own licenses. Each recipe names its model’s license, and tensorless info MODEL shows it before anything is downloaded.
| Model | License |
|---|---|
| Qwen3.8-27B-PN3 | Apache-2.0, from Qwen/Qwen3.8-27B |
| Qwen3.8-27B-UD-IQ3_XXS-Q3Floor | Apache-2.0, from Qwen/Qwen3.8-27B and unsloth/Qwen3.8-27B-GGUF |
Read a model’s card on Hugging Face for its full terms.