Fine-tuning APIs: who still lets you train a model, how, and at what price

Offered with no wind-down noted: Amazon Bedrock, Microsoft Foundry, Together AI, Fireworks AI, Alibaba Model Studio, Nebius Token Factory.

• Together AILoRA supervised fine-tuning, Preference fine-tuning (DPO), LoRA and Full-parameter supervised fine-tuning

  The pricing page shows two unlabeled tables under LoRA and Full fine-tuning tabs. We read the first as LoRA and the second as full fine-tuning by tab order, because the second table's models match the docs' full fine-tuning list.

• Alibaba Model StudioContinual pre-training (CPT) and Direct preference optimization (DPO)

  These model lists come from the table under the second region tab of the overview. We attribute that table to North China 2 (Beijing) by tab order; the page text does not name the region beside it.

• Microsoft FoundryManaged supervised fine-tuning and Managed reinforcement fine-tuning

  The pricing page renders prices for a default region and does not state which one. The Regional, Global and Developer prices shown are for that unstated region.

• Google Vertex AIVertex AI supervised fine-tuning (Gemini)

  Status recorded as “available (per-model GA vs preview not stated in page text)” is inferred from the docs rather than stated. Limitation-table anchors for Gemini 3.5 Flash and Gemini 3.1 Flash-Lite are named '...-public-preview', suggesting those two are Preview; not stated in visible text.

• Amazon BedrockSupervised fine-tuning

  Status recorded as “GA (inferred)” is inferred from the docs rather than stated.

SFTSupervised fine-tuning

You show the model example prompts with the answers you want, and it learns to answer like that. The default method almost everywhere.

DPODirect preference optimization (preference tuning)

You give pairs of answers to the same prompt, one preferred and one not, and the model learns to lean toward the preferred kind. Good for tone and style.

RFTReinforcement fine-tuning

The model attempts tasks and a grader (your code, or another model) scores each attempt; training pushes it toward higher-scoring answers. Often billed by training hour, plus the grader's tokens.

Vision SFTVision fine-tuning

Supervised fine-tuning where the examples include images, so the model gets better at a visual task.

DistillationDistillation

A large teacher model writes answers that are used to fine-tune a smaller, cheaper student model. You pay for the teacher's calls as well as the training.

LoRALoRA (adapter) fine-tuning

Trains a small set of extra weights, an adapter, on top of the frozen base model. Cheaper and faster than full fine-tuning; the adapter is served alongside the base model.

FullFull-parameter fine-tuning

Updates every weight in the model. Costs more than LoRA and produces a complete copy of the model.

CPTContinued pre-training

Keeps training the model on a large body of raw, unlabeled text from your domain, so it absorbs vocabulary and facts before any instruction tuning.