Learn
Toy → Inferix client
M0 toy → general LLM → domain SLMs → register on Inferix. Same physics, different jobs.
The tiny GPT taught the loop. Owned product models reuse it at useful scale, then Inferix operates them. Do not skip the train. Do not wait on Inferix to start M1.
Order (do not invert)
M0 nanoGPT (done when ckpt.pt reloads)
M0b dissect a small open model (specimen)
M1 general-llm LoRA/SFT on 3B-class instruct → local /generate
M2 slm-support narrow CX data, same train/eval/serve
M3 slm-apiheal contract classify; heal tools denied
then agents
then Inferix: register the same *-vN as clientsSame physics, different jobs
| Family | Job | Later RouteIQ path |
|---|---|---|
| M0 tiny GPT | Literacy + your first checkpoint | Not registered |
| general-llm | Broad owned reasoning | General / strong-owned |
| slm-support | Cheap CX / FAQ | Cheap CX |
| slm-apiheal | Cheap classify / summarize | Cheap classify; no heal |
When Inferix is ready
- Register endpoints in
inferix.yaml— Owned models guide - RouteIQ picks cheap vs general vs provider
- LensAI / TraceForge label
model_id - DriftWatch vs teacher; FineForge promote / rollback
Quality bar
No mock trainer, no stub model for dashboards. Hardware tight → smaller real model, not a fake train. Gaps vs Claude are expected and useful — that gap is the drift story later.