Skip to content
Best LLM Fine-Tuning CompaniesGet a quote
Pages

Independent scorecard · eight vendors · buyer side

Best LLM Fine-Tuning Companies

Eight fine-tuning vendors scored from their own pricing and documentation pages. Tell us the job, and we name the vendors whose published terms fit it before anyone sees your brief.

Get a quoteOpen the scorecard

1 Brief

The job, in plain words

Task
Pull 14 fields from claim forms
Base
Open-weight, 8B class
Examples
About 4,000 reviewed pairs
Runs in
Our AWS account

2 Shortlist

Vendors whose terms fit

  • Cloud platformAWS Bedrock
  • Cloud platformGoogle Vertex AI
  • Open-model cloudFireworks AI

3 Pilot

A small run on your test set

Prompting only71%
Tuned adapter86%

Target agreed in the brief: 85%

4 Contract

Terms you can hold

  • Adapter weights exportable to you
  • Test set and grader stay yours
  • Training data deleted, in writing
  • Serving price basis stated

Invented example brief and pilot scores. The vendor names are real; nothing here is a quote or a test we ran.

The scorecard

Four vendors publish enough to score 6 or more

Each vendor is marked on five criteria: two points where its own pages publish the answer in full, one for a partial answer, none where they say nothing. Together AI leads on 7 of 10, with Google Vertex AI, Fireworks AI and AWS Bedrock on 6.

  • Published pricing
  • Model choice
  • Evaluation support
  • Data handling
  • Deployment

See the reason behind every square

  1. Together AIOpen-model cloud, API
    7/10
  2. Google Vertex AICloud platform
    6/10
  3. Fireworks AIOpen-model cloud, API
    6/10
  4. AWS BedrockCloud platform
    6/10
  5. Azure OpenAICloud platform
    4/10
  6. AnyscaleCompute platform (Ray)
    1/10
  7. Prime IntellectRL training and environments
    1/10
  8. OpenPipeAgent RL training (CoreWeave)
    0/10

Independent: no company named here owns, sponsors or has endorsed this site. Vendors on this list can owe this site a fee if an introduction it makes ends in a signed contract. Squares, order and wording are set without regard to that.

Not scored

OpenAI fine-tuning is closed to new teams

OpenAI is phasing out fine-tuning on its own platform, according to its pricing page, and no longer takes new customers for it. Existing customers may keep creating jobs for the coming months, and their tuned models stay callable until each base model is retired.

A new project cannot start there, so it is left off the scorecard rather than marked down.

“OpenAI is winding down the fine-tuning platform.”

OpenAI, API pricing page

  • New sign-upsClosed
  • Existing customersFor the coming months
  • On this scorecardNot scored

Four kinds of fine-tuning company, and what each is for

The name on the contract matters less than the kind of vendor. Each kind settles where training runs, who writes the code and how the bill is built. Detail and sources for every vendor sit in the vendor notes.

Open-model clouds

Per-token training on open-weight models, served from their endpoint

Together AI has the widest open-model menu on the list, from sub-1B Qwen3.5 models to DeepSeek and GLM, with a minimum charge per job. Fireworks AI prices managed training by base-model size and adds nothing on top to serve a tuned model: you pay deployment time in GPU-seconds.

$0.34Together AI, Llama 3.1 8B SFT, per 1M tokens
$4.00Together AI minimum charge per job, most models
$0Fireworks AI surcharge to serve a tuned model; GPU time is still billed
Together AIFireworks AI

Cloud platforms

The tuned model stays in your cloud account

Google Vertex AI tunes Gemini and open models from one price page. AWS Bedrock prices reinforcement fine-tuning by the hour and charges a monthly storage fee per custom model. Azure OpenAI lists hourly hosting alongside training.

Google Vertex AIAWS BedrockAzure OpenAI

Compute platform

Your training code, their clusters

Anyscale runs post-training on Ray, as LoRA adapters or full-weight updates with SFT, DPO or RLVR. Pricing is by quote.

Anyscale

RL training

When a grader matters more than a file of examples

Prime Intellect hosts an open hub for RL and evaluation environments. CoreWeave announced it was acquiring OpenPipe, an agent RL specialist we have yet to score.

Prime IntellectOpenPipe

Data handling

Blank for all eight

None of the pricing and documentation pages we read state how long training data is kept, whether it trains other models or how it is deleted. Ask each vendor for its data processing terms in writing.

Settle these eight points before asking for a quote

Vendors quote faster and more honestly once they are written down. None of them needs you to send data.

  1. 01

    The task in one sentence

    What the tuned model must do better than the base model, stated so an outsider could judge it.

  2. 02

    A baseline and a target

    Today's score with prompting or retrieval, and the score that would justify tuning.

  3. 03

    Base model family

    Closed API, open-weight, or must run on your hardware. This alone rules out half the list.

  4. 04

    Examples on hand

    Roughly how many reviewed input and output pairs exist, and who can label more.

  5. 05

    Where the model must run

    The vendor's endpoint, your cloud account or on premises, plus expected monthly traffic.

  6. 06

    Data rules

    Retention limits, regions and whether the vendor may use training data for anything else.

  7. 07

    The review a vendor must pass

    A SOC 2 report, a data processing agreement or your own questionnaire.

  8. 08

    Who owns the result

    Tuned weights or adapters, the evaluation set and the grader, if one is used.

Have most of these? We will tell you which vendors' published terms line up with them.

Get a quote

What fine-tuning vendors charge to train

The bill counts your dataset's tokens once per epoch, and serving the tuned model is a second, separate bill. Units differ, so compare like with like. All prices are on the scorecard.

  • $0.34Together AI's published fine-tuning price, Llama 3.1 8B SFT, per 1M tokens
  • $0.50Fireworks AI, LoRA SFT on base models up to 16B, per 1M training tokens
  • $0.67Google Vertex AI, Llama 3.1 8B SFT, per 1M training tokens
  • $8.00Fireworks AI, dedicated H100 training compute, per GPU-hour
  • $80.00AWS Bedrock, reinforcement fine-tuning of Qwen3 32B, per training hour

Prices as each vendor prints them. Not offers. Azure OpenAI figures did not load for us, and only part of the Bedrock page was recorded.

Independent: no company named here owns, sponsors or has endorsed this site. Vendors on this list can owe this site a fee if an introduction it makes ends in a signed contract. Squares, order and wording are set without regard to that.

Where these facts come from (7)
  1. Pricing, Fireworks AI
  2. Pricing, Together AI
  3. Fine-tuning pricing, Together AI docs
  4. Generative AI on Vertex AI pricing, Google Cloud
  5. Bedrock pricing, AWS (only part of the page was recorded)
  6. Azure OpenAI Service pricing, Microsoft (figures did not load)
  7. API pricing, OpenAI

Buyers' questions about fine-tuning companies

Still unsure which kind of vendor fits? Describe the job and we will name a few.

Which company is best for LLM fine-tuning?

It depends on the base model and where the model must run. On published information alone, Together AI scores highest of the eight (7 of 10), with Google Vertex AI, Fireworks AI and AWS Bedrock next on 6. The full scorecard shows why each square is filled.

What is the best platform for fine-tuning an LLM?

For open-weight models served from someone else's endpoint, Together AI and Fireworks AI publish per-token training prices and how a tuned model is served. If the model has to live in a cloud account you already run, Google Vertex AI and AWS Bedrock tune Gemini, Llama, Qwen and other models there. Teams that want to run their own training code on managed clusters look at a compute platform such as Anyscale.

Can a new team still fine-tune through OpenAI?

No. On OpenAI's API pricing page, the fine-tuning platform is described as winding down, with sign-ups closed to new users. Teams already on it may keep creating training jobs for the coming months. A new project has to start somewhere else, which is why OpenAI carries no score here.

What does LLM fine-tuning cost?

Published training prices start at $0.34 per million tokens (Together AI, Llama 3.1 8B, SFT) and $0.50 (Fireworks AI, LoRA SFT up to 16B). Billed tokens are your dataset multiplied by epochs, and serving the tuned model is a separate, recurring cost.

Should we hire a consultancy or use a platform?

Use a platform if you already have clean examples and a test set. Hire help when the hard part is the data, the evaluation or the deployment. Our post on consultancy or provider tools sets out the signals.

Who owns the fine-tuned model?

Whoever the contract names, so write it down before work starts. Ask who owns the tuned weights or adapters and whether you can export them; if the base model is open-weight, its license still applies to the result. The evaluation set and grader should belong to you, since every future retrain is judged against them. Our note on what a statement of work should include covers the clauses.

Do you test the vendors?

No. We read each vendor's own published pages and score what they state. We do not run training jobs, benchmark models or audit security. The scoring method lists exactly what counts.

Get a quote

Start with a ZIP code, add an email, then tap through what you want tuned. We read it, check which vendors' published terms fit, and write back with names before anyone else sees your brief.

This is an introduction service. The vendor you are introduced to may pay it a fee if you sign work together; the team asking is never charged.

Step 1 of 3

Where is the company based?

Vendors differ on the entities and data regions they accept, so location trims the list.

Leave out data, samples, prompts and credentials. A plain description of the project is enough to shortlist vendors.

Get a quote