Find releases of OPEN-WEIGHT language, vision-language, audio or embedding models published in the LAST 21 DAYS. Each finding must be ONE model repository or release page on huggingface.co, modelscope.cn, GitHub, or the lab's own site, for ONE model family. MUST-HAVES: the weights are actually downloadable, not merely announced or waitlisted; the LICENCE is stated (Apache 2.0, MIT, Llama Community, Gemma, Qwen, a custom research-only licence, or non-commercial); the parameter count or the family's sizes are stated; and the page carries a date inside the 21-day window. LEAD WITH THE LICENCE, because that is what decides whether the model is usable at all: a research-only or non-commercial licence is a different product from Apache 2.0 however good the benchmarks are. Then the parameter count, the context window, and the modality. Skip re-uploads, GGUF and AWQ quantisations, LoRA adapters and fine-tunes of an existing release, merges, 'best open models' listicles, leaderboard pages, and anything outside the window. A quantised copy of a model already sent is not a new release. Lead with the licence, the model, the sizes, and the lab.