Custom Models
We train your model.
On your data and our open-source engine, served where your data lives.
- 2.53× training vs. Unsloth
- 7.0× serving vs. llama.cpp
- Apache 2.0 engine
From your data to a served model.
The same engine runs every step, on your GPUs or ours.
Data
Your documents, chats and records, versioned in the hub with every checkpoint.
Train
Full fine-tuning, LoRA or QLoRA, then GRPO, DPO or distillation.
Evaluate
HumanEval, MBPP, terminal-bench, SWE-bench or your own suite, before anything ships.
Serve
On the same engine, on your GPUs or ours. List it on llm.buzz at your own price.
We train our own first.
Our speech models and our engine are the proof: published, measured, and open to check.
A model trained on a team’s own chats.
Two fine-tunes scored against a big model on 20 questions the training chats never asked. Filmed on a demo business.
Where it runs.
Your model goes where your data goes.
Our EU cloud
Hosted in the EU. Start in minutes.
Your cloud
Your account, your region, your keys.
Your hardware
On your servers, managed by your team.
Air-gapped
No connection out. Models included.
Tell us what your model should do.
A partner near you, or our own engineers.