Fine-tune a model

Fine-tuning adapts a model's behavior to your data — teaching it your domain, style, or task — without changing its size. It's the fast, low-cost counterpart to growing a model: same model, new behavior. It works on any model size.

Fine-tune vs. grow

	Fine-tune	Train / grow
What it does	adapts behavior on the existing model	adds capacity, then trains it
Model size	unchanged	larger
Use when	you want the model to behave differently	you want a smarter / bigger model
Speed & cost	fastest, cheapest	more involved

Both produce a new model in your account and can be followed by contraction.

Run one

bash

curl https://console.axomlabs.ai/api/jobs \
  -H "Authorization: Bearer $AXOM_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "finetune",
    "idempotency_key": "ft-2026-06-04-a",
    "source_model_id": "<your-model-id>",
    "dataset_id": "<dataset-id>",
    "hyperparams": { "lora_rank": 32, "steps": 2000, "output_name": "support-tone-v1" }
  }'

Fine-tune a model from your account with source_model_id, or a shared base model with base_model_id. You get back a queued job; monitor it and download the result like any run.

Settings

Setting	Default	What it controls
`lora_rank`	`32`	Adapter capacity — higher learns more nuance, costs a little more.
`steps`	`2000`	How long it trains. More steps = more adaptation (diminishing past ~3000).
`output_name`	auto	Name for the new model.
(advanced) `lr`, `seq_len`	`1e-4`, `256`	Standard tuning knobs.

Strength — pick a level

Level	`lora_rank` · `steps`	Use it for
Light	16 · 1000	Quick, cheap behavior nudges.
Balanced (recommended)	32 · 2000	The quality-per-dollar sweet spot.
Deep	64 · 3000	Maximum adaptation on harder tasks.

Not sure? Call POST /training/recommend with type: "finetune" for a tailored starting point.

Pricing

Fine-tuning is billed on training tokens only — there's no growth fee (the model doesn't get bigger), so it's substantially cheaper than growing. Tokens = steps × sequence_length. Estimate it:

bash

curl https://console.axomlabs.ai/api/pricing/estimate \
  -H "Authorization: Bearer $AXOM_KEY" -H "Content-Type: application/json" \
  -d '{ "type": "finetune", "params_after": 2360000000, "steps": 2000 }'

See Pricing for the full model.

What you get back

A new, adapted model (same size as the input) plus a result summary with the final loss, tokens processed, and cost. The original is never modified.

Fine-tune a model ​

Fine-tune vs. grow ​

Run one ​

Settings ​

Strength — pick a level ​

Pricing ​

What you get back ​

Next ​