What is the best platform for fine-tuning LLMs?
Published by Best LLM Fine-Tuning Companies
The best platform for fine-tuning an LLM is the one that supports the base model you want to tune, with the method you need, and serves the result in a way your traffic and security review accept. For open models, Together AI and Fireworks AI publish the most complete prices. For Gemini, Google Vertex AI is the platform on our list, and teams already committed to AWS or Azure usually start with Bedrock or Azure OpenAI.
We have not tested any of these platforms. Everything below comes from each vendor’s own pricing and documentation pages, scored the way our method page describes.
Best LLM fine-tuning platforms, by what they publish
| Platform | Kind | What it tunes | Published training price | Serving |
|---|---|---|---|---|
| Together AI | Open-model cloud, API | Llama, Qwen, Gemma, Mixtral, gpt-oss, DeepSeek, Kimi, GLM | From $0.34 per 1M tokens (Llama 3.1 8B, SFT); $4 to $60 minimum per job | Dedicated endpoints, per minute |
| Fireworks AI | Open-model cloud, API | Four size tiers, up to 16B to over 300B; LoRA or full | LoRA SFT $0.50 to $10.00 per 1M training tokens; full is double | No serving surcharge; deployments per GPU-second |
| Google Vertex AI | Cloud platform | Gemini 3.5, 3.1, 2.5; Gemma 3, Llama, Qwen 3 | Gemini per 1,000 tokens (3.5 Flash $0.01); Llama 3.1 8B $0.67 per 1M | Tuned Gemini 3+ at base price globally, 1.5x regionally |
| AWS Bedrock | Cloud platform | gpt-oss-20b, Qwen3 32B, Llama models | RFT $80.00 per training hour; Llama 2 70B $7.99 per 1M | Per token or provisioned per hour, plus $1.95 per model per month |
| Azure OpenAI | Cloud platform | OpenAI models only (o4-mini, GPT-4.1, GPT-4o families) | Listed, but figures did not load for us | Hourly hosting charge per tuned model |
| Anyscale | Compute platform (Ray) | Your own training code: SFT, DPO or RLVR, LoRA or full | Quote only | Not covered in the docs |
Vendors on this list can owe this site a fee if an introduction it makes ends in a signed contract. Rows, order and wording are set without regard to that. Independent: no company named here owns, sponsors or has endorsed this site.
What is the best fine-tuning API provider?
If you want to upload a file and get a tuned model back, the API-style platforms are Together AI and Fireworks AI for open models, and Vertex AI for Gemini. Together AI lists the widest open-model menu we recorded, from sub-1B Qwen3.5 models up to Qwen3.5 397B A17B, DeepSeek and GLM. Fireworks AI has the simplest price sheet: one rate per model-size tier, with LoRA and full tuning side by side.
Which fine-tuning API is easiest to use?
We do not test usability, so we cannot rank it. What the published pages do show is how predictable the bill is. Fireworks AI prices by model size alone. Together AI adds validation tokens to the billed count and sets a minimum per job, but refunds failed jobs in full, per its docs. Vertex AI mixes units: Gemini per thousand tokens, open models per million. Try a small run on two platforms with the same data before you commit.
What is the best serverless fine-tuning platform?
“Serverless” usually means you never manage GPUs for training or serving. On published terms, per-token serving of a tuned model exists on Vertex AI (tuned Gemini) and on Bedrock for its reinforcement-tuned models. Fireworks AI and Together AI serve tuned models on deployments or dedicated endpoints billed by time, which suits steady traffic but bills while idle.
Best open source LLM for fine-tuning
The platforms above tune the open families most teams start with: Llama, Qwen, Gemma, Mixtral, gpt-oss and DeepSeek. The best one is the smallest model that passes your evaluation, because smaller models are cheaper to train and much cheaper to serve. Pick two or three sizes, tune each on the same data with LoRA, and compare on your own test set. Check the base model’s license for your use, since it still applies to the tuned result.
Best companies to fine-tune an open source LLM for you
If you would rather not run the project yourself, the choice is between a platform plus your own team and a firm that does the data, evaluation and training for you. Our post on whether to hire a fine-tuning consultancy or use a platform sets out when each makes sense, and the list of best LLM fine-tuning companies separates platforms from firms that do the work.
How to choose in four questions
- Which base model families must you tune, and at what size?
- Which method: supervised, preference (DPO) or reinforcement against a grader?
- Where must the tuned model run, and is per-hour hosting acceptable at your traffic?
- What will the platform say in writing about retention and training use of your data? None of the pricing pages we scored states it.
The scorecard scores each platform on those points from its published pages. Choose the platform that tunes your model; the rest is price.
Sources for this post (8)
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.
Brief received.
- We read it against the published terms, model lists and deployment options of the vendors on our scorecard.
- We write back with the vendors that fit and why, by name. Nothing has gone to any of them yet.
- Want an introduction? Tell us which vendor. Your brief and contact details move only after that reply.