# Fine-tuning platforms

> Fine-tuning teaches a model your format, tone or task from your own examples. There are three ways in. Together AI, Fireworks, Google Cloud and AWS run managed jobs: upload a dataset, pay per training token, call the tuned model. Tinker sits in the middle: you write the training loop in Python, it runs the GPUs. Unsloth and Axolotl are free, open-source trainers for hardware you control. OpenAI's fine-tuning is closed to new users and winding down, so it isn't here. Training tokens are your dataset times the number of epochs, so a small clean dataset beats a big messy one on the bill too.

7 tools, in the order they appear in the Digidle Tools directory. This is not a ranking.

- Page: https://digidle.com/best/fine-tuning-platforms/
- All best-of lists: https://digidle.com/best/

## Tools

- [Together AI](https://digidle.com/tools/together-ai/): Serverless API for open models such as gpt-oss, GLM and Kimi, OpenAI-compatible and billed per token, with batch, provisioned throughput, dedicated endpoints, fine-tuning (LoRA or full, SFT and DPO, priced per training token) and GPU clusters when you outgrow shared capacity. Image generation models too, such as FLUX.2, GPT Image 2 and Nano Banana Pro, priced per image. Video generation models as well, such as Seedance 2.5, HappyHorse and Hailuo 02. Official site: https://www.together.ai. Alternatives: https://digidle.com/alternatives/together-ai/.
- [Fireworks AI](https://digidle.com/tools/fireworks-ai/): Inference platform for open models: per-token serverless with Standard, Priority and Fast modes, batch at half the serverless price, on-demand GPUs billed per second, and LoRA or full fine-tuning (SFT, DPO) with tuned models served at the base price. Serverless embedding models too, from Qwen3 Embedding 8B to small BERT-size models. OpenAI-compatible. Official site: https://fireworks.ai. Alternatives: https://digidle.com/alternatives/fireworks-ai/.
- [Google Cloud Gemini Tuning](https://digidle.com/tools/gemini-tuning/): Google Cloud's managed fine-tuning for Gemini in Gemini Enterprise Agent Platform, formerly Vertex AI: supervised fine-tuning on Gemini 3.5 Flash, 3.5 Flash-Lite, 3.1 Flash-Lite and the 2.5 models, preference tuning on 2.5 Flash and Flash-Lite, and RL fine-tuning with up to 16 reward functions on 3.5 Flash and 3.1 Flash-Lite. From $1.50 per million training tokens on 2.5 Flash-Lite. Gemma, Llama and Qwen open models tune too. Official site: https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/tuning. Alternatives: https://digidle.com/alternatives/gemini-tuning/.
- [Amazon Bedrock Model Customization](https://digidle.com/tools/bedrock-model-customization/): Amazon Bedrock's model customization: supervised fine-tuning, reinforcement fine-tuning with reward functions in AWS Lambda, and distillation from a large teacher model into a smaller one. Billed per training token times epochs, from $1 per million on Nova Micro, plus $1.95 per model a month for storage. Parameter-efficient custom models run on demand at the base model's price. Official site: https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html. Alternatives: https://digidle.com/alternatives/bedrock-model-customization/.
- [Tinker (Thinking Machines)](https://digidle.com/tools/tinker/): Thinking Machines' fine-tuning API for open-weight models from 4B to 1T parameters: write the training loop in Python on your own machine with forward_backward, optim_step and sample, and Tinker runs it on its GPUs. LoRA only, for SFT, RL, DPO and distillation. Pay per token, from $0.396 per million training tokens, and download any saved checkpoint. Official site: https://thinkingmachines.ai/tinker/. Alternatives: https://digidle.com/alternatives/tinker/.
- [Unsloth](https://digidle.com/tools/unsloth/): Open-source fine-tuning library and desktop app for your own GPU or a free Colab notebook: LoRA, QLoRA, full fine-tuning, GRPO, DPO and FP8 for LLMs, vision, TTS and embedding models, which Unsloth says train 2x faster with 70% less VRAM. Exports to GGUF. Core Apache-2.0, Studio UI AGPL-3.0; Windows, Linux and macOS. Official site: https://unsloth.ai. Alternatives: https://digidle.com/alternatives/unsloth/.
- [Axolotl](https://digidle.com/tools/axolotl/): Free, open-source (Apache-2.0) fine-tuning framework: one YAML config covers dataset prep, training, evaluation and quantization. Full fine-tuning, LoRA, QLoRA, DPO, KTO, ORPO and GRPO on NVIDIA or AMD GPUs, multi-GPU with FSDP2 or DeepSpeed and multi-node with Torchrun or Ray. Official site: https://axolotl.ai. Alternatives: https://digidle.com/alternatives/axolotl/.

## At a glance

| Tool | What it is | Methods | Training price | Models and deploy |
| --- | --- | --- | --- | --- |
| [Together AI](https://digidle.com/tools/together-ai/) | Inference cloud with managed fine-tuning jobs, plus Post-Training on GPUs you reserve. | LoRA or full fine-tuning. SFT and DPO, with vision, function-calling and reasoning data. Post-Training adds GRPO. | LoRA SFT from $0.34 per 1M training tokens (Llama 3.1 8B, small Qwen3.5), Llama 3.3 70B $2.03. DPO costs about 2.5x. Minimum charge per job from $4. | Open models or your own Hugging Face model. Serve on a dedicated endpoint or download the weights. |
| [Fireworks AI](https://digidle.com/tools/fireworks-ai/) | Inference platform with managed fine-tuning and a Training API. | SFT, DPO and ORPO, as LoRA or full-parameter. Datasets in OpenAI chat JSONL. | LoRA SFT per 1M training tokens: $0.50 up to 16B, $3 to 80B, $6 to 300B, $10 above. Full fine-tuning and DPO cost double. | Open models. Tuned models run on on-demand deployments, multi-LoRA included, at the base model's price. |
| [Google Cloud Gemini Tuning](https://digidle.com/tools/gemini-tuning/) | Managed tuning in Gemini Enterprise Agent Platform, formerly Vertex AI. | Supervised fine-tuning on Gemini 3.5 Flash, 3.5 Flash-Lite, 3.1 Flash-Lite and 2.5. Preference tuning on 2.5 Flash and Flash-Lite. RL fine-tuning with reward functions on 3.5 Flash and 3.1 Flash-Lite. | Per 1M training tokens: 2.5 Flash-Lite $1.50, 3.5 and 3.1 Flash-Lite $3, 2.5 Flash $5, 3.5 Flash $10, 2.5 Pro $25. Gemma 3 1B $0.47. | Gemini plus open models such as Gemma, Llama and Qwen. Tuned Gemini 3 and later cost the same as the base model in the global region, 1.5x in others. |
| [Amazon Bedrock Model Customization](https://digidle.com/tools/bedrock-model-customization/) | Model customization inside Amazon Bedrock. | Supervised fine-tuning, reinforcement fine-tuning with reward functions in AWS Lambda, distillation from a teacher model. | Per 1M training tokens: Nova Micro $1, Nova Lite $2, Nova 2 Lite $3.78, Nova Pro $8. RFT on Nova 2 Lite $80 an hour. Storage $1.95 per model a month. US East. | Amazon Nova plus open models such as Llama, gpt-oss and Qwen. Parameter-efficient custom models run on demand at the base price; full-rank ones need Provisioned Throughput. |
| [Tinker (Thinking Machines)](https://digidle.com/tools/tinker/) | Thinking Machines' training API: your loop in Python, their GPUs. | LoRA only. SFT, RL (GRPO, PPO), DPO, distillation or your own loss, through forward_backward, optim_step and sample. | Per 1M tokens, by model: training from $0.396 (GPT-OSS-20B), Qwen3-8B $0.44, Kimi-K2.6 $4.84. Sampling and prefill billed apart. Storage $0.10 per GB a month. | Open-weight models from 4B to 1T parameters, dense and MoE. Download any saved checkpoint. |
| [Unsloth](https://digidle.com/tools/unsloth/) | Open-source trainer and desktop app for your own GPU. | LoRA, QLoRA, full fine-tuning, pretraining, GRPO, DPO, FP8. LLMs, vision, TTS, embedding and diffusion models. | Free. Free Colab notebooks to start, then your own GPU. | Core Apache-2.0, Studio UI AGPL-3.0. Windows, Linux, WSL and macOS; NVIDIA, AMD and Intel GPUs. Exports to GGUF. |
| [Axolotl](https://digidle.com/tools/axolotl/) | Open-source fine-tuning framework run from one YAML config. | Full fine-tuning, LoRA, QLoRA, QAT, DPO, IPO, KTO, ORPO, GRPO, reward models. Multimodal too. | Free. You pay for the GPUs. | Apache-2.0. NVIDIA or AMD GPUs, multi-GPU with FSDP2 or DeepSpeed, multi-node with Torchrun or Ray. Docker images and PyPI. |

Checked 2026-11-02 on vendor pages: [Together AI pricing](https://www.together.ai/pricing), [Together AI fine-tuning docs](https://docs.together.ai/docs/fine-tuning-overview), [Fireworks pricing](https://fireworks.ai/pricing), [Fireworks fine-tuning docs](https://docs.fireworks.ai/fine-tuning/fine-tuning-models), [Gemini Enterprise Agent Platform tuning docs](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/tuning), [Google Cloud generative AI pricing](https://cloud.google.com/vertex-ai/generative-ai/pricing), [Amazon Bedrock custom models docs](https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html), [Amazon Bedrock pricing](https://aws.amazon.com/bedrock/pricing/), [AWS price list for Bedrock, US East](https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/us-east-1/index.json), [Tinker](https://thinkingmachines.ai/tinker/), [Tinker pricing](https://thinkingmachines.ai/tinker/pricing/), [Tinker docs](https://tinker-docs.thinkingmachines.ai/), [Unsloth on GitHub](https://github.com/unslothai/unsloth), [Axolotl on GitHub](https://github.com/axolotl-ai-cloud/axolotl).

## Pick it if

- [Together AI](https://digidle.com/tools/together-ai/): You want LoRA or full fine-tuning on open models, a posted price per model, and the weights to take home.
- [Fireworks AI](https://digidle.com/tools/fireworks-ai/): You'll serve the tuned model on Fireworks and want it billed like the base model.
- [Google Cloud Gemini Tuning](https://digidle.com/tools/gemini-tuning/): You want a tuned Gemini, or RL fine-tuning with graded rewards, without running a trainer.
- [Amazon Bedrock Model Customization](https://digidle.com/tools/bedrock-model-customization/): You're on AWS and want to tune Nova or distill a big model into a small one on your AWS bill.
- [Tinker (Thinking Machines)](https://digidle.com/tools/tinker/): You're doing RL or research and want to own the loss and the loop, not the GPUs.
- [Unsloth](https://digidle.com/tools/unsloth/): You have one GPU, or a free Colab, and want to fine-tune on it today.
- [Axolotl](https://digidle.com/tools/axolotl/): You run your own cluster and want one config file from dataset to multi-node training.

---
Source: Digidle Tools (https://digidle.com/). Full index: https://digidle.com/llms.txt