dwain-barnes

I fine-tune small language models, quantise them to GGUF and build voice agents that run entirely on local hardware. All of it ships to Hugging Face. Eryri Labs is one person and four RTX 3090s, and the name comes from the mountains outside the window.

Model description

The models I make are small on purpose. A 30B that runs in 24GB of VRAM is useful to a security team that can't send its detection logic to a cloud API. An 8B legislation model is useful to someone who wants answers without a subscription. That's the niche: capable models for people who need the weights in the building.

Recent releases include Glimmer-Sentry-30B, a detection engineering assistant that translates between Sigma, KQL, SPL and YARA (every training pair generated with the official sigma-cli converter, not another LLM), and Caelum-G4-38B, a sparse mixture-of-experts merge that outscores all four models it was built from.

There is also a Reachy Mini on the desk. It tracks whoever is speaking in a group conversation and turns to face them. Everything runs locally; nothing leaves the room.

Intended use

Building small, private, local models for people who need them.

Training details

Every release taught the next one something. Fine-tunes led to merges, merges to distillation, distillation to a mixture-of-experts, and lately a 30B trained overnight in seven and a half hours, one epoch, on a single 3090.

Evaluation

these numbers come straight from the Hugging Face API and update themselves

Limitations

• Based roughly 5,000 miles from the Bay Area. Meetings stay short.
• Will explain GGUF quantisation at parties, unprompted.
• The robot has better posture than the engineer.

How to get this model

You can't. But the language models are all open, the code is on GitHub, and the inbox is friendly.