2B parameters. 4 citations. 0 bytes left the machine.
We've been pushing ManualIQ hard on modest hardware.
This week: gemma4:e2b — a 2 billion parameter model — running completely local on a 5-year-old M1 MacBook with 8GB RAM.
One question: what is the crosswind limitation? Four answers came back. Each one cited:
Praetor 600 AFM — Page 57. Page 357. Page 376. Page 459.
No hallucination. No cloud. No API key. Not a single byte left the machine.
Here's what months of building taught us:
The model was never the hard part. Finding the right 6 pages inside a 400-page aviation manual — that's the hard part. Once ManualIQ retrieves them, even a 2B model reads and formats the answer correctly.
Retrieval over inference. Every time. This is why it works for the nursing student who can't afford a RAM upgrade. The pilot whose licensed manuals can't go to ChatGPT. The small law firm that needs answers from their own documents, not the internet's best guess.
Your document is already the answer. ManualIQ just finds it.
v3.1 live now at cavokdesigns.com
Early adopter pricing — code PH29 at checkout.
#LocalAI #RAG #SovereignComputing #Aviation #ManualIQ #homeofficedogstevie
Replies