Fireworks Fine-Tuning
Fireworks AIFireworks Fine-Tuning runs managed supervised, DPO and ORPO preference, and reinforcement fine-tuning on open-weight models such as Llama, Mistral, Qwen and DeepSeek, producing LoRA adapters that can warm-start from a prior checkpoint. A Training API adds custom Python training loops on Fireworks-managed GPUs, with full-parameter training on dedicated capacity. Trained models are served only from Fireworks on-demand deployments, either merged into the base model or as multi-LoRA adapters sharing hardware.
Openness
1 high confidence- source
- closed
- training-pipeline
- closed
- runs-on
- Fireworks-cloud-only
- license
- Proprietary(managed service)
- base-weights
- open-models(varies)
The managed fine-tuning service and its training pipeline are proprietary and run only on Fireworks' cloud, with no source published. Fireworks does publish an open-source reinforcement-fine-tuning toolkit, reward-kit, but it has not been confirmed as part of this product.
- https://docs.fireworks.ai/fine-tuning/fine-tuning-models recorded 2026-09-23
The docs describe SFT, DPO/ORPO, and reinforcement fine-tuning jobs run via Fireworks' own dashboard, CLI (firectl), and API, with the tuned LoRA model deployed to an on-demand (dedicated) deployment, "the only supported method for serving fine-tuned models"; no self-hosting or downloadable engine is described anywhere on the page.
Adoption
not assessedFireworks publishes no usage figures for its fine-tuning service, such as jobs run, customers tuning or models fine-tuned. The platform serves substantial inference volume overall, but none of it is specific to fine-tuning, so the map leaves this product unscored rather than borrowing a figure from the wider platform.
- https://docs.fireworks.ai/fine-tuning/fine-tuning-models recorded 2026-09-28
Searched the body for customers, developers, trillion, jobs run, monthly active and users: no usage figure for fine-tuning is published; the only "million" is a dataset limit of "3 million dataset examples".
Capability
4 medium confidenceFireworks offers managed supervised, preference and reinforcement fine-tuning that produces LoRA adapters, plus a Training API for custom training loops, with full-parameter training on dedicated GPUs by request. It runs only on Fireworks' cloud and serves trained adapters only from dedicated deployments, which puts it level with managed services such as Tinker and Together rather than with Megatron-LM and the large-scale training frameworks.
- https://docs.fireworks.ai/fine-tuning/deploying-loras.md recorded 2026-09-28
The page states "Serverless deployment is not supported for LoRA models" and compares live merge, which "Matches the base model (no overhead)", with multi-LoRA, which has "Some overhead per request due to dynamic adapter application".
- https://docs.fireworks.ai/fine-tuning/fine-tuning-models recorded 2026-09-28
The page documents supervised, DPO/ORPO and reinforcement fine-tuning jobs, continued training with "--warm-start-from" instead of a base model, and says training a model and deploying it to an on-demand (dedicated) deployment "is the only supported method for serving trained models".
- https://docs.fireworks.ai/fine-tuning/training-api/dedicated.md recorded 2026-09-28
The dedicated training page says to "Use cookbook recipes with LoRA or full-parameter configurations supported by the selected training shape" and lists SFT, DPO, ORPO, RL, distillation and custom loops.
- https://docs.fireworks.ai/fine-tuning/training-api/introduction.md recorded 2026-09-28
The Training API overview describes "custom training loops with full Python control over objectives" on Fireworks-managed GPUs, and says Dedicated Training is for when "You need full-parameter training".
Verified 2026-09-23