Small models,
trained in the open.
We build language models efficient enough to run on hardware you own, and release the weights, code and datasets behind them under Apache-2.0.
Hugging Face, October 2026. Every number is public: huggingface.co/empero-ai
Qwen3.8 Distill
One teacher, several students: Qwen3.8-2.4T distilled into compact reasoning models with a 262K context and native tool use.
Model cards ↗Qwythos
Our reasoning family that checks its own work, with a 1M-token context on consumer hardware.
Model cards ↗Qwen3.8-27B-Ridge
A mixed-precision GGUF that fits a 27B model on a single consumer GPU while keeping quality close to BF16.
Model card ↗Every release ships day-one GGUF builds for llama.cpp, Ollama, LM Studio and vLLM. Our models also power Abacus, a native terminal coding agent that runs on any OpenAI-compatible endpoint, including your own machine.
From distillation to models of our own.
Our next models are trained from scratch: highly sparse, and designed to be everyday assistants on the devices people already have. We announce models when the weights are ready to download, not before.
Follow the build.
A short letter every other Tuesday. No hype, cancel from any line.
Scale with us.
We're looking for long-term compute partners to grow our models with us, step by step, from our first from-scratch models toward the trillion-parameter range.
We're especially interested in cost-efficient hardware: AMD Instinct, NPUs and other accelerators. Paid, discounted or sponsored: we're open to whatever works long term for both sides.
- Partners and investors. If you build, host or back accelerators, we'd like to hear from you.
- Developers. Issues and PRs are read by the people who trained the models.
- Enterprises. Apache-2.0, self-hosted and auditable down to the weights.