Start supervised fine-tuning: firectl sftj create --base-model <BASE_MODEL_ID> --dataset <DATASET_ID> --output-model <FINE_TUNED_MODEL_ID>. Use --warm-start-from <PRIOR_MODEL_ID> to continue training an existing adapter.
Poll with firectl sftj get <JOB_ID> until it completes. The result is registered as accounts/<ACCOUNT_ID>/models/<FINE_TUNED_MODEL_ID>.
Option A - live merge, for a single adapter with no per-request overhead: firectl deployment create accounts/<ACCOUNT_ID>/models/<FINE_TUNED_MODEL_ID>. The LoRA weights are merged into the base model at deploy time.
Query a live-merge deployment: POST https://api.fireworks.ai/inference/v1/chat/completions with model set to accounts/<ACCOUNT_ID>/models/<FINE_TUNED_MODEL_ID> and header 'Authorization: Bearer <FIREWORKS_API_KEY>'.
Option B - multi-LoRA, to serve many adapters on one shared base: create the base deployment with firectl deployment create accounts/fireworks/models/<BASE_MODEL_ID> --enable-addons, then attach the adapter with firectl load-lora <FINE_TUNED_MODEL_ID> --deployment <DEPLOYMENT_ID>.
Query a multi-LoRA deployment with model set to accounts/<ACCOUNT_ID>/models/<FINE_TUNED_MODEL_ID>#accounts/<ACCOUNT_ID>/deployments/<DEPLOYMENT_ID>.
Benchmark both paths before committing: live merge minimizes latency for one adapter, multi-LoRA maximizes density when serving many.
Verify the base model supports LoRA deployment before you spend on a training run - not every base model does.
Official documentation: https://docs.fireworks.ai/fine-tuning/fine-tuning-models | https://docs.fireworks.ai/fine-tuning/deploying-loras | https://docs.fireworks.ai/api-reference/create-supervised-fine-tuning-job
Known gotchas
Fine-tuned LoRA models cannot run on serverless. They require a dedicated on-demand deployment, which bills per GPU-hour even when idle.
Multi-LoRA adds roughly 10-30 percent time-to-first-token overhead versus live merge, because adapters are applied per request instead of pre-merged.
Multi-LoRA requires a BF16 base deployment. You cannot combine --enable-addons with FP8 or FP4 quantized deployment shapes.
The default LoRA addon quota is 100 per account; high-tenancy adapter serving needs a quota increase.
Some base models cannot host LoRA adapters at all. Confirm compatibility first - discovering it after a fine-tuning run wastes both time and spend.
The pure-REST equivalents of the firectl dataset upload and SFT job creation commands are documented less precisely than the CLI. Confirm exact paths at the create-supervised-fine-tuning-job reference before scripting against REST instead of firectl.
Give your agent this knowledge — and 16,000+ more routes
One MCP install gives any agent live access to the full route map across 5,800+ domains, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp
Need this verified for your stack — or a route we don't have yet?