Recent analysis of six LLMs claimed to be "free" reveals that none currently meet the four essential freedoms defined by the GNU Project, primarily because their training data includes non-free or attribution-encumbered material. Richard Stallman and contributors argue that for an LLM to be truly free, its training dataset—the functional equivalent of source code—must consist exclusively of freely licensed data, a standard not met by models like Apertus, Moxin, OLMo, Marin, DaVinci, SmolLM, or OpenCoder due to their reliance on datasets like Common Crawl or The Stack, which contain copyrighted works. Consequently, until a genuinely free LLM is developed, Emacs packages relying on them should not be recommended, and users remain ultimately responsible for copyright issues in generated output regardless of the model's licensing status.
VIDRAFT Releases Aether-7B-5Attn: A Fully Open-Source MoE LLM with Five Heterogeneous Attention Mechanisms