Skip to content

Best of · Fine-tuning

The best LLM fine-tuning platforms in 2026

The best LLM fine-tuning platforms in 2026 are Together AI for managed per-token fine-tuning of open models with downloadable weights, Unsloth for free, fast training on your own GPU, Tinker for writing your own training loop on managed GPUs, W&B Training (OpenPipe's successor) for agent reinforcement learning, and Vertex AI or Amazon Bedrock if you're tied to those clouds. Kojen ranks lower for training itself (LoRA on two 8B models, early access) but is the place to test whether you need to fine-tune at all.

The market shifted in 2026: OpenAI closed fine-tuning to new users, OpenPipe's hosted platform shut down, Fireworks paused managed tuning and Predibase was folded into Rubrik.

By Will SchulzUpdated 29 sources

The list

  1. 01Together AI

    www.together.ai

    Best for: managed fine-tuning of open models, priced per token

    Published per-token prices (SFT from $0.34 per million tokens for 8B-class models), LoRA and full fine-tuning with SFT and DPO, GRPO through Post-Training, downloadable weights, LLM-judged evaluations and dedicated endpoints.

    Pricing: SFT from $0.34 per 1M tokens for 8B-class models up to $2.03 for Llama 3.3 70B (DPO costs more; $4 minimum per job). Post-Training on B200: $11/GPU-hour. Source · Open source: no

    Strengths

    • Transparent per-token prices on a wide range of open-weight models.
    • You can download the tuned weights or adapter.
    • Evaluations, fine-tuning, RL and serving on one platform.

    Limits

    • Dedicated endpoints bill while running, even when idle.
    • RL is a separate GPU-minute product that needs credit for four hours up front.
  2. 02Unsloth

    unsloth.ai

    Best for: fast, free fine-tuning on your own GPU

    The fastest, lowest-memory open-source trainer, covering LoRA, QLoRA, full fine-tuning, DPO and GRPO, with GGUF export and local serving. You run the hardware.

    Pricing: Open source: free. Pro and Enterprise tiers (multi-GPU, multi-node) are listed without public prices. Source · Open source: yes (Apache-2.0 core; some parts AGPL-3.0)

    Strengths

    • The fastest, least memory-hungry way to fine-tune on one GPU, with a huge community.
    • Covers everything from QLoRA to full fine-tuning and GRPO reinforcement learning.
    • Free and runs entirely on hardware you control.

    Limits

    • You supply, run and pay for the GPUs.
    • No evaluations, red-teaming or hosted deployment; paid-tier prices aren't published.
  3. Best for: custom training loops without managing GPUs

    Write SFT, DPO or RL in Python and Tinker runs it as LoRA on its GPUs, across very large models, priced per token. Limited production serving.

    Pricing: Per 1M tokens, by model: e.g. Qwen3-8B training $0.44; larger models cost more. Checkpoint storage $0.10/GB-month. Source · Open source: no

    Strengths

    • Full control of the training algorithm without managing GPUs.
    • Very large models available, priced per token.
    • Adapters export to standard formats.

    Limits

    • You write the training code.
    • LoRA only, and no general production serving.
  4. 04OpenPipe (now ART + W&B Training)

    github.com/OpenPipe/ART

    Best for: reinforcement learning for agents

    OpenPipe's ART library (free) and its managed successor, W&B Training on CoreWeave at $2.70 per GPU-hour, train LoRA adapters with GRPO and the RULER judge, then serve them on W&B Inference.

    Pricing: ART: free, Apache-2.0. Successor CoreWeave serverless SFT and RL: $2.70 per GPU-hour of active training, with rollouts and storage billed separately. Source · Open source: yes (Apache-2.0 (ART))

    Strengths

    • ART is one of the best open-source libraries for training agents with reinforcement learning.
    • RULER removes the need to hand-write reward functions.
    • Managed GPU-hour pricing on a large cloud through W&B Training.

    Limits

    • The no-code fine-tuning SaaS, with its managed evals and DPO, no longer exists.
    • Managed use now means a CoreWeave / W&B account.
  5. Best for: tuning Gemini on Google Cloud

    The only way to tune Gemini, inexpensive per token, with an evaluation service and prompt optimizer alongside. Gemini weights can't be exported.

    Pricing: Gemini SFT from $0.0015 to $0.025 per 1K training tokens by model; open models e.g. Llama 3.1 8B $0.67 per 1M tokens. Source · Open source: no

    Strengths

    • The only way to tune Gemini, and inexpensive per token.
    • Evaluation and prompt optimization in the same cloud.
    • Enterprise controls of Google Cloud.

    Limits

    • Tied to Google Cloud.
    • Tuned Gemini weights can't be exported.
  6. Best for: teams that must stay inside AWS

    Supervised fine-tuning for Nova, Claude 3 Haiku and Llama, reinforcement fine-tuning with an LLM judge as grader, and Bedrock Evaluations. A short model list and costly RFT hours ($80 per hour).

    Pricing: RFT $80 per training hour for gpt-oss-20B and Qwen3 32B; $1.95 per custom model per month storage; per-token Nova tuning prices vary. Source · Open source: no

    Strengths

    • Data stays inside AWS with enterprise compliance.
    • Reinforcement fine-tuning with an LLM judge as the grader.
    • Can fine-tune Claude 3 Haiku and Amazon Nova.

    Limits

    • A short list of tunable models.
    • Expensive RFT hours and provisioned throughput.
  7. Best for: the widest choice of methods at multi-node scale

    A config-driven open-source framework supporting full, LoRA, QLoRA, DPO, KTO, ORPO and GRPO with FSDP2 and DeepSpeed. Entirely do-it-yourself.

    Pricing: Free, Apache-2.0. You pay for your own compute. Source · Open source: yes (Apache-2.0)

    Strengths

    • The broadest menu of training methods.
    • Scales from one GPU to multi-node clusters.
    • Reproducible YAML configs.

    Limits

    • Entirely do-it-yourself infrastructure.
    • No evaluation or deployment.
  8. 08Kojenour product

    kojen.ai

    Best for: deciding whether to fine-tune, and testing the result

    Kojen's text LoRA trains a real adapter on Qwen3-8B or Llama-3.1-8B-Instruct on the cheapest GPU in stock across Modal, Runpod, Vast.ai and Lambda, billed by the second, but it is early access and adapters can't be served yet. Its strength is the cheaper levers first (prompt tuning, model search) and evaluation with confidence intervals around every version. We make Kojen.

    Pricing: Free to start, no card needed; then pay as you go. GPU training (LoRA) is paid in credits billed by the second at the GPU's live price. Team and volume pricing by email. Source · Open source: no

    Strengths

    • Measures and improves in one loop: evaluations, red-teaming, prompt tuning, model search and deployment share the same project and test cases.
    • Honest statistics by default: every rate has a confidence interval, tuned prompts must beat the current one in a paired test, and pass@k / pass^k come from several samples per case.
    • Agent-native: the full loop is exposed over an MCP server and a REST API with an OpenAPI spec, with the same 24 tools as the in-app agent.

    Limits

    • Young product: weight training is early access on 8B models, LoRA adapters can't be served yet, and full fine-tuning and RLHF are previews.
    • No tracing SDK for apps running elsewhere, so it is not a replacement for an observability platform.
    • Closed source and hosted only; no self-hosting.

Side-by-side comparison

"Partial" means the feature exists with real limits, given in the note. "Not verified" means we couldn't confirm it from the vendor's own material.

The best LLM fine-tuning platforms in 2026: features and pricing, checked November 2026
Checked Nov 15, 2026Together AIUnslothTinkerOpenPipe (now ART + W&B Training)Google Vertex AI tuningKojen
What it isFine-tuning

Together AI (independent)

Fine-tuning

Unsloth AI (independent)

Fine-tuning

Thinking Machines Lab

Fine-tuning

CoreWeave (acquisition announced September 3, 2025)

Fine-tuning

Google Cloud

Training dojo

Kojen (that's us)

PricingSFT from $0.34 per 1M tokens for 8B-class models up to $2.03 for Llama 3.3 70B (DPO costs more; $4 minimum per job). Post-Training on B200: $11/GPU-hour. SourceOpen source: free. Pro and Enterprise tiers (multi-GPU, multi-node) are listed without public prices. SourcePer 1M tokens, by model: e.g. Qwen3-8B training $0.44; larger models cost more. Checkpoint storage $0.10/GB-month. SourceART: free, Apache-2.0. Successor CoreWeave serverless SFT and RL: $2.70 per GPU-hour of active training, with rollouts and storage billed separately. SourceGemini SFT from $0.0015 to $0.025 per 1K training tokens by model; open models e.g. Llama 3.1 8B $0.67 per 1M tokens. SourceFree to start, no card needed; then pay as you go. GPU training (LoRA) is paid in credits billed by the second at the GPU's live price. Team and volume pricing by email. Source
Open sourceNoYes, Apache-2.0 core; some parts AGPL-3.0 RepoNoYes, Apache-2.0 (ART) RepoNoNo
Fine-tuning / LoRA

Trains model weights or adapters.

Yes

LoRA and full; SFT, DPO; GRPO via Post-Training

Yes

LoRA, QLoRA, full FT, DPO, GRPO; vision and TTS

Yes

LoRA with SFT, DPO and RL; you write the loop

Yes

GRPO into LoRA adapters (ART); SFT via W&B Training

Yes

SFT, preference and RL tuning

Partial

Text LoRA on Qwen3-8B / Llama-3.1-8B, early access; adapters can't be served yet; full fine-tuning and RLHF are previews

Model endpoints

Serves a model or tuned version behind an API you call in production.

Yes

Dedicated endpoints (billed while running)

Partial

Local OpenAI-compatible serving in Studio; no hosted endpoints

Partial

Sampling client; adapters export to Hugging Face PEFT

Yes

Through W&B Inference

Yes

Managed endpoints

Yes

OpenAI-compatible, SSE streaming, tool calls forwarded

Evaluations on datasets

Run a model or app over a set of test cases and score the results.

Yes

Classify, score and compare evaluations

No
No
No

The old platform's managed evals were not migrated

Yes

Gen AI evaluation service

Yes

Test cases scored for accuracy, hallucination, bias and latency

LLM-as-judge scoring

A model grades outputs against a rubric or reference.

Yes
No
No
Yes

RULER, an LLM judge that produces RL rewards

Yes

Autorater with rubrics

Yes

Judge kept out of the target's model family in simulations

Prompt optimization

Automatically rewrites instructions or examples to raise a metric.

Not verified
No
No
No
Yes

Zero-shot and data-driven prompt optimizer

Yes

Prompt tuning plus search across 11 hosted models

Confidence intervals

Rates reported with an interval or a significance test, not just a point estimate.

Not verified
No
No
No
Not verified
Yes

Wilson and bootstrap intervals on every rate

Official MCP server

An MCP server from the vendor that lets agents drive the product.

Not verified
No

Local models can be used from MCP clients; no Unsloth MCP server

Not verified
No

Has an example that trains models to use an MCP server's tools

Not verified
Yes

ai.kojen/kojen in the MCP Registry, OAuth sign-in

Self-host

You can run it on your own infrastructure.

Partial

Weights downloadable; the service itself is hosted

Yes

Runs on your own GPUs

No
Yes

ART runs on your own GPU

No

No Gemini weight export

No

Hosted only

What changed in 2026

OpenAI stopped new organizations from creating fine-tuning jobs on May 7, 2026, and active customers lose job creation on January 6, 2027; tuned models keep serving until their base models retire. OpenPipe, bought by CoreWeave, stopped hosted training and inference on July 30, 2026 and moved them to W&B Training. Fireworks AI paused managed RL on October 1, 2026 and now steers customers to its code-first Training API. Predibase, acquired by Rubrik in 2025, no longer has a standalone site.

The lesson for choosing a platform is portability: prefer services that let you download the weights or adapter, and keep your training data in a format you control.

Before you fine-tune

Fine-tuning is the expensive, slow lever. Many tasks reach the same quality with a better system prompt or a different base model, which costs minutes, not GPU-hours. Measure the gap with a held-out test set first, and fine-tune only when prompt tuning and model choice can't close it. Kojen's free calculators estimate the GPU memory and cost if you do.

How we made this list

  • We compared supported methods (LoRA, QLoRA, full fine-tuning, DPO, RL), model choice, published prices, weight export, and whether evaluation and serving come with it.
  • Order reflects how well each platform does the job of fine-tuning in November 2026; services that closed or paused managed fine-tuning are covered in the text, not ranked.
  • Every feature claim links to the vendor's own pricing page, docs or repository, checked on November 15, 2026. Where we couldn't confirm a feature, the table says "Not verified".
  • We make Kojen. We include it where it does the job and say where others are stronger; it is not placed first by default.
  • No vendor paid to be included, and there are no affiliate links.

Frequently asked questions

Can I still fine-tune OpenAI models?
Only if your organization is an active existing fine-tuning customer, and only until January 6, 2027, per OpenAI's deprecations page. New organizations lost access on May 7, 2026.
What is the cheapest way to fine-tune an LLM?
On hardware you already own, Unsloth or Axolotl cost nothing beyond electricity. Managed, Together AI's SFT for 8B-class models starts at $0.34 per million training tokens, and Tinker trains Qwen3-8B at $0.44 per million.
Which platforms let me download my fine-tuned weights?
Together AI, Tinker (adapters to Hugging Face PEFT format), and every self-run tool (Unsloth, Axolotl, TRL, ART). OpenAI and Vertex AI's Gemini tuning do not.
What happened to Fireworks fine-tuning?
Fireworks paused managed reinforcement fine-tuning on October 1, 2026, and its docs say managed SFT and DPO no longer accept new jobs; its code-first Training API remains.

Keep reading

Sources

Checked November 2026.

  1. [1]Pricing. Together AI.
  2. [2]Fine-tuning overview. Together AI docs.
  3. [3]AI evaluations. Together AI docs.
  4. [4]Deploying a fine-tuned model. Together AI docs.
  5. [5]Post-training pricing. Together AI docs.
  6. [6]unslothai/unsloth. GitHub.
  7. [7]Unsloth documentation. Unsloth.
  8. [8]Pricing. Unsloth.
  9. [9]Tinker: general availability and vision input. Thinking Machines.
  10. [10]Models & pricing. Tinker docs.
  11. [11]Tinker pricing. Thinking Machines.
  12. [12]CoreWeave to acquire OpenPipe, leader in reinforcement learning. CoreWeave.
  13. [13]CoreWeave acquires agent-training startup OpenPipe. TechCrunch.
  14. [14]OpenPipe (platform migration notice). OpenPipe.
  15. [15]OpenPipe/ART. GitHub.
  16. [16]Post-training pricing. CoreWeave.
  17. [17]Generative AI pricing. Google Cloud.
  18. [18]Gen AI evaluation service. Google Cloud docs.
  19. [19]Custom models. AWS docs.
  20. [20]Supported models for customization. AWS docs.
  21. [21]Reinforcement fine-tuning. AWS docs.
  22. [22]Bedrock pricing. AWS.
  23. [23]axolotl-ai-cloud/axolotl. GitHub.
  24. [24]Kojen for Agents: REST API & MCP Server. Kojen.
  25. [25]llms.txt. Kojen.
  26. [26]Pricing. Kojen.
  27. [27]Rubrik to acquire Predibase to accelerate agentic AI adoption. Rubrik.
  28. [28]predibase/lorax (archived). GitHub.
  29. [29]Deprecations: Evals platform. OpenAI docs.

Try it on your own test cases.

Run an LLM-judged evaluation with confidence intervals, a red-team pass and a prompt-tuning run on your agent. Free to start, no card needed, then pay as you go.

Start training free