LLM Fine-Tuning Services
Pages
Get a quote

Fine-tuning service providers

Fine-tuning providers fall into two groups: platforms where your team runs the training (Fireworks AI, Together AI, Google Vertex AI, AWS Bedrock, Azure OpenAI, Anyscale) and specialists who focus on one part of the work, such as reinforcement learning for agents or building the training data. The tables show what each publishes about methods, pricing and hosting.

Last verified , each line against the company's own pricing or product page.

No company in these tables owns, sponsors or has approved this page. Nothing here was tested by us: each cell restates what the company itself published, checked on 7 November 2026, and "Not published" marks anything a company does not state.

Which platforms let your team fine-tune a model?

Self-serve platforms with published fine-tuning offers, in alphabetical order.

ProviderWhat it offersPublished pricing basisHosting the tuned model
AnyscaleCompute platform (Ray)Ray-based post-training with LLaMA-Factory, SkyRL or Ray Train. Methods in its docs include supervised tuning, DPO and RLVR, using adapters or full weights.Not published for fine-tuningNot published
AWS BedrockCloud model platformCustomization of hosted models, including reinforcement tuning; its price list also has a distillation section.Pricing page only partly captured.Per 1M training tokens (Llama 2) or $80.00 per training hour (reinforcement fine-tuning)Per token, or provisioned model units per hour; $1.95 per model per month storage
Azure OpenAICloud model platformTuning for the GPT-4.1 family (full, mini, nano), GPT-4o and GPT-4o-mini, plus reinforcement tuning for o4-mini.Token-based training (hourly for o4-mini); amounts not capturedCharged hourly while a tuned model is deployed
Fireworks AIInference and training platformHosted tuning of open models with adapters or full weights, supervised or DPO, priced in tiers by model size. GPUs can also be rented hourly for training.$0.50 to $10.00 per 1M training tokens for LoRA SFT, by model sizeNo serving surcharge; deployments metered in GPU-seconds
Google Vertex AICloud model platformTuning for Gemini (supervised on all listed, preference or reinforcement on some) and for open families such as Llama, Gemma and Qwen.Gemini per 1,000 training tokens; open models per 1MGemini 3 onward: base rate on the global endpoint, 1.5x regionally
Together AIInference and fine-tuning cloudSupervised and DPO tuning across a long list of open models, charged by the token with a floor price on each job.From $0.34 per 1M tokens (SFT, Llama 3.1 8B); minimum $4.00 to $60.00 per jobPer-minute dedicated endpoints, rate set by hardware

Disclosure: some firms in this list can pay us a referral fee when one of our introductions turns into paid work. Inclusion, order and wording are decided without regard to that.

Who handles reinforcement learning or the training data?

Firms focused on one stage of the project rather than general fine-tuning, in alphabetical order.

ProviderWhat it offersPublished pricingRecent change
OpenPipe (CoreWeave)Reinforcement learning for agentsReinforcement learning tooling for training agents, and ART, an open-source library.Not publishedNow owned by CoreWeave (deal announced 3 September 2025).
Prime IntellectOpen RL training and environmentsRuns an open Environments Hub where people publish RL and eval environments; also offers prime-rl (open source) and Lab, a hosted trainer.Not publishedHub opened on 27 August 2025.
SuperAnnotateData platform for training setsSoftware for producing the labeled sets a tuning project needs, for training and for evaluation alike.Three tiers (Starter, Pro, Enterprise), amounts unlistedNone noted

Disclosure: some firms in this list can pay us a referral fee when one of our introductions turns into paid work. Inclusion, order and wording are decided without regard to that.

What about OpenAI fine-tuning?

OpenAI's price list now carries a notice that fine-tuning is being retired and is shut to new customers; current customers can launch jobs for some months more. For a project starting now, look elsewhere. Source: the OpenAI API price list, seen 7 Nov 2026. For teams with jobs still running, the final published rates sit on our prices page.

What should you ask a provider before signing?

  1. Which base models and methods? LoRA or full fine-tuning, SFT, DPO, reinforcement fine-tuning. Check the exact model versions you plan to use.
  2. Where does training run, and where does serving run? Shared cloud, dedicated hardware or your own account. This decides most of the security review.
  3. Who owns the tuned weights, and can you export them? Some platforms only serve tuned models in place.
  4. Is your data retained or used for anything else? Get the answer in the contract, not a blog post.
  5. What does serving cost at your volume? Ask for the billing unit (per token, per GPU-second, per hour) and a worked example at your expected traffic.
  6. How will you evaluate the result? A provider who does not ask about your evaluation set early is a warning sign.
Sources behind this section

Questions about choosing a provider

What is the difference between a fine-tuning platform and a fine-tuning service?

A platform gives your engineers the tools and GPUs: you prepare the data, launch the job and run the evaluation. A service firm does some or all of that work for you, from building the training set to handing over a tuned, tested model. Most platforms on this page publish per-token prices; service work is priced per project.

Which provider is cheapest for fine-tuning?

It depends on the model and method. For small open models, the published rates on this page start at $0.28 per million training tokens (Llama 3.2 1B on Vertex AI) and $0.34 (Llama 3.1 8B SFT on Together AI, with a $4.00 minimum per job). Serving costs after training often matter more. See the price tables.

Can a provider fine-tune a model inside our own cloud account?

Some can, but it is rarely on the pricing page. Ask directly whether training and serving can run in your account or virtual private cloud, who holds the tuned weights afterward, and whether your data is kept or used for anything else.

Do you rank or test these providers?

No. We do not test, benchmark or audit providers. This page lists what each one publishes about itself, in alphabetical order within each group, with the source and the date we read it.

Why is OpenAI not in the main table?

Its own price list announces that fine-tuning is being retired and no longer accepts new customers. A table of options for new projects should not include it.

Get a quote

Only three things are needed: the company's ZIP code, where to send our answer, and a short sketch of the project in taps. The answer names providers whose published offers line up with it.

Keep data out of this form: no samples, records, weights or keys. Describing the task in words gives us enough to work with.

This site is an introduction service, and a provider it introduces you to may pay it a fee. Your details stay with us until you approve a named provider.

1 of 3: location

Where is the company based?
Get a quote