An HVAC & electrical sizing co-pilot that runs on your own box.
SLM Advisor is a 3-billion-parameter language model fine-tuned for residential HVAC, electrical, and MEP sizing work. It loads locally on a laptop or shop server, answers without per-token API fees, and ships with LoRA adapters you can re-export to GGUF or MLX for your own toolchain.
What you actually get
Five things this addon does that a vanilla chat model doesn't, out of the box.
Sizing-aware reasoning
Trained on residential load calc prompts, NEC ampacity tables, generator kW selection, and duct/pipe sizing heuristics — so it answers in the formats you already use.
Electrical + HVAC bilingual
Switches between AC/DC ampacity, voltage drop, OCPD selection, and refrigerant/tonnage/CFM problems without losing units.
Local-only inference
Runs through llama.cpp, Ollama, or TabbyAPI on your hardware. No calls leave the box. No token bills. No rate limits.
Inspectable adapters
LoRA adapters ship in safetensors and are re-exportable to GGUF (llama.cpp) and MLX (Apple Silicon). Merge into a full checkpoint when you want.
Drop-in to your stack
OpenAI-compatible API mode via TabbyAPI, plus a raw llama.cpp CLI. Plug into existing field tools, terminals, or your own RAG layer.
Honest licensing
Base model is Qwen2.5-3B (Apache-2.0). Fine-tuned weights and LoRA adapters are released under MIT for the personal & shop tiers — commercial terms on the Firm license.
Technical stack (honest version)
No marketing fluff — here's exactly what you're getting under the hood.
| Layer | Choice | Notes |
|---|---|---|
| Base model | Qwen2.5-3B | Apache-2.0 upstream; small enough to run on shop hardware |
| Fine-tune | Unsloth LoRA | 2× faster than vanilla HF, fits on a 12 GB card |
| Training data | HVAC + electrical corpora | Built in-house, no scraped customer data |
| Inference | llama.cpp / Ollama / TabbyAPI | Your choice — local on your hardware |
| Export format | safetensors → GGUF / MLX | Recipes in README |
| License | MIT (shop tiers) | Base keeps Apache-2.0; Firm adds commercial clause |
There is no hosted API. There is no telemetry collection. There is no third-party model in the loop.
Three licenses. No subscriptions.
Pay once, download the weights, keep using them. Renew only if you want the next training data drop.
For solo techs running the model on a laptop or shop PC.
- Qwen2.5-3B base + LoRA adapter (safetensors)
- Personal & single-seat commercial use
- GGUF export recipe (Q4_K_M, Q5_K_M, Q8_0)
- MLX export recipe for Apple Silicon
- README + system-prompt starter kit
- Email support, 72h response window
For working contractors running a small team or job-site server.
- Everything in Starter
- Up to 5 seats at one physical site
- TabbyAPI OpenAI-compatible server config
- Pre-merged GGUF + MLX checkpoints (no merging on your end)
- Field-prompt pack: 30 sizing workflows
- Quarterly LoRA refresh drops for 12 months
- Email + Discord channel, 24h response window
For MEP firms that need broad internal use and integration rights.
- Everything in Pro
- Unlimited seats within one legal entity
- Commercial redistribution right inside your own product
- SOC-friendly inference profile & logging recipe
- Custom LoRA fine-tune slot (1 request per year)
- Priority support, 8h business-hours response
- Optional on-call video walkthrough (1 hr)
Pick the tier that matches your seat count.
You'll get a Whop checkout link with the exact deliverables listed. Downloads are signed; weights are checksum-verified.
Questions buyers actually ask
Straight answers, no marketing fog.
What hardware do I need to run it?
The Q4_K_M GGUF is roughly 2.0 GB and runs on 8 GB unified memory (M1/M2/M3 Mac mini, Steam Deck, modern Mini-PCs). Q8_0 wants around 4 GB. A 6 GB consumer GPU is plenty for the full safetensors build via TabbyAPI.
Is any of my data sent to a server?
No. The model files run locally through llama.cpp, Ollama, or TabbyAPI on your own hardware. The only network traffic is the Whop download and optional email support.
Can I re-export the model to another format?
Yes. LoRA adapters ship in standard safetensors; the README includes one-command recipes to export to GGUF (llama.cpp) and MLX (Apple Silicon). You can also request the pre-merged checkpoints on the Pro tier.
What is the difference between the tiers?
Starter gets you the weights + recipes for personal/shop use. Pro adds pre-merged checkpoints, server configs, a 30-prompt field pack, and quarterly refresher updates for a year. Firm adds commercial redistribution rights, unlimited seats, and priority support.
Do I need an internet connection to use it?
Only once — to download the model bundle from Whop after purchase. After that, everything (the model, the UI, the API) runs fully offline. Nothing is phoned home.