Independent lab,
open by default.

We build language models efficient enough to run on hardware you own, and we release the weights, the code and the datasets that produced them. Latest: a full-parameter distill of Qwen3.8-2.4T into 9B / 4B / 2B, and a GDN-aware 27B Ridge GGUF at 11.7 GiB. The open-weight Qwythos family — 9B and 27B, 1M-token context — has passed one million downloads on Hugging Face. Everything we ship is Apache-2.0.

We own the stack end to end: rethink generates the reasoning traces, SFTSuite orders them into curricula, microverse searches architectures before we commit a training run, and FTPO fixes failure modes without a retrain. Two next-generation in-house models are in pretraining; we announce models when the weights are ready to download, not before.