Models
The model family
Each model is a fine-tune of the same base (Qwen3.6-35B-A3B, Apache-2.0) for one vertical, with a point-of-use guardrail baked into the weights.
The family
Call a model by its slug (the model field). Every slug below is live on Hugging Face for self-hosting and on the hosted API. The catalog has the full card — taglines, evals, and install commands — for each.
| Slug | Niche | Hugging Face |
|---|---|---|
legal-intake | Law-Firm Intake Coordinator | flywheel-ai/legal-intake |
healthcare-frontdesk | Healthcare Front Desk | flywheel-ai/healthcare-frontdesk |
automotive | Auto Repair & Service Ops | flywheel-ai/automotive |
home-services | Home Services (HVAC · Plumbing · Electrical) | flywheel-ai/home-services |
beauty-wellness | Beauty & Wellness (Salon / Spa / Med-Spa) | flywheel-ai/beauty-wellness |
restaurant | Restaurant & Hospitality | flywheel-ai/restaurant |
fitness | Fitness | flywheel-ai/fitness |
construction | Construction & Trades | flywheel-ai/construction |
agency-ops | AI-Automation Agency Operator | flywheel-ai/agency-ops |
real-estate | Real Estate Agent Ops | flywheel-ai/real-estate |
Base & sizes
All models share one base — Qwen3.6-35B-A3B, an Apache-2.0 mixture-of-experts model — so they have identical runtime characteristics and you can swap niches without re-plumbing. Each ships in two builds:
- GGUF (Q4_K_M) — ~20 GB. Laptop- and CPU-friendly; runs in llama.cpp.
- bf16 safetensors — ~65 GB. Full precision for GPU serving with vLLM.
Picking a model
Choose the model whose vertical matches your business — a gym uses fitness, an auto shop uses automotive. Each is tuned for the language, tasks, and guardrails of that trade, so a niche model out-answers a general model of the same size on its home turf. If no niche fits, a private model can be trained for yours.
Versioning & guardrails
Models are versioned (v1.0, …) and a new version ships only when it measurably beats the last. With consent, real usage trains the next version — the flywheel. Every model also carries a point-of-use guardrail (its output is decision support, not professional advice) baked into both the weights and the runner config, so it travels with the model wherever you run it.