Model components / Fine-tuning code

Fireworks Fine-Tuning

Managed LoRA fine-tuning for popular open-weight models (Llama, Mistral, Qwen, DeepSeek) with served checkpoints at base-model inference rates and tight integration to Fireworks' high-throughput inference engine. Picked when latency-sensitive production serving matters as much as the fine-tune itself; Fireworks consistently benchmarks at the low end of TTFT for hosted open-weight models. Pricing from $0.50/M training tokens for models up to 16B.

Fireworks AI managed fine-tuning (docs live June 2026). Methods: SFT, DPO, RFT (reinforcement fine-tuning). Output produces LoRA models; warm-start continued training supported; fine-tuned models served on on-demand (dedicated) deployments only. Proprietary managed service over open base models.

Openness

1 high confidence

1.0

source: closed
training-pipeline: closed
runs-on: Fireworks-cloud-only
license: Proprietary(managed service)
base weights: open-models(varies)

The managed FT service/pipeline is proprietary and runs only on Fireworks' cloud; no source. (Fireworks' separate reward-kit RFT toolkit is OSS but was not confirmed as part of this product this run.)

https://docs.fireworks.ai/fine-tuning/fine-tuning-models recorded 2026-06-04
managed FT service (SFT/DPO/RFT); LoRA outputs; dedicated-deployment serving; proprietary cloud

Adoption

not assessed

No disclosed standalone usage figure for the Fireworks fine-tuning feature located this run (jobs run / customers tuning). Fireworks serves notable inference volume, but no honest signal specific to the FT SKU was found; declining to assign a level.

https://docs.fireworks.ai/fine-tuning/fine-tuning-models recorded 2026-06-04
managed FT feature; no usage figures disclosed

Capability

4 medium confidence

4.0

Strong managed method coverage (SFT+DPO+RFT, incl. reinforcement FT) over open base models. Black-box (LoRA-output oriented, limited scale/precision control), no MLPerf-Training basis, so capped below OSS scale definers; scored 4, on par with the other RFT-capable managed services. Slightly narrower demonstrated scale than Together (no confirmed 100B+ full-FT claim), but matching RFT depth keeps it at 4.

https://docs.fireworks.ai/fine-tuning/fine-tuning-models recorded 2026-06-04
SFT, DPO, RFT; LoRA outputs + warm-start

Unchanged since 2026-06-09 (last edited, not re-checked)