Open-Weight Models Under MIT License

SAVRN's Take

We read every license the same way: what it lets a facility do with the weights, and what it wants back. MIT wants very little. It permits commercial use, modification and redistribution, and the one condition is that the copyright notice and permission notice stay with the files. An operator can quantize a model, fine-tune it on customer data, fold it into a product and sell inference on it, and compliance comes down to keeping those notices in the bundle.

What carries MIT on our hub skews small. Of the 425 models under it, feature extraction leads the tasks at 54 models and text generation has 41. bge-small-en-v1.5 from the Beijing Academy of Artificial Intelligence tops downloads at 64,516,396 a month with 33M parameters, a 512-token context and 0.1 GB at 16-bit; the cheapest listing is one MI300X at $1.85 an hour, and it rides on the same card as whatever generation model you already serve. bge-m3 follows at 38,175,398 with an 8,194-token context for longer passages, then xlm-roberta-base at 21,479,919 and gpt2 at 15,439,333.

Microsoft leads the publishers with 30 models, then Moritz Borrett-Laurer at 18, AI at Meta at 17, the Beijing Academy at 10 and DeepSeek at 8. Before you commit, check that the license on the model card matches the files you pulled from the publisher; a mismatch there is the one way this short license gets complicated.

SAVRN Research, 2026-09-18