Google Cloud, Amazon Bedrock, Microsoft Foundry, and Together AI let you fine-tune a model without managing infrastructure. AutoScientist, Adaption’s system for automating the research loop behind model training, goes further by setting the training recipe and adding data. Unsloth Studio, TRL, and Axolotl run on your own hardware.
Key Takeaways
- Managed cloud platforms from Google Cloud, Amazon Bedrock, and Microsoft Foundry run fine-tuning inside a cloud account many enterprises already use, and each publishes commitments about how your training data is handled.
- Together AI runs training for open models and lets you download the trained weights, so you can run the finished model somewhere else.
- AutoScientist, Adaption’s system for self-improving and automating the research loop behind model training, co-optimizes your data and the training recipe, and the trained model’s weights export to your own infrastructure.
- Open-source tools such as Unsloth Studio, TRL, and Axolotl keep your data and the resulting model on hardware you control.
What Is LLM Fine-Tuning?
LLM fine-tuning is additional training of a pretrained large language model (LLM) on your own examples, so the model learns your domain’s language, output format, or behavior. Most models are built for the average case, and prompt engineering works around that by rewording requests. Fine-tuning changes the model itself, which pays off in three ways:
- Consistency. The model applies your terminology and returns the structure your application expects, more reliably across requests.
- Accurate tool use. The model calls tools correctly because it has been trained on examples of tools in use.
- Lower cost. Training replaces long prompts, so token spend per call drops, and a smaller trained model can replace a large proprietary one for a specific task.
What Tools Allow You To Fine-Tune A Model?
Eight tools allow you to fine-tune a model: Google Cloud, Amazon Bedrock, Microsoft Foundry, Together AI, AutoScientist, Unsloth Studio, TRL, and Axolotl. They fall into four paths based on how much work they take on: a cloud you already use, a specialist platform, automated training, or an in-house build.
- Inside the cloud you already use: Google Cloud, Amazon Bedrock, and Microsoft Foundry run training for you, and you choose the method and settings.
- A specialist platform: Together AI does the same for open-weight models and lets you download the result.
- Automated training: AutoScientist sets the training recipe and adds data, and the trained model’s weights export to your own infrastructure.
- Build in-house: Unsloth Studio, TRL, and Axolotl run on hardware you control, and your team sets everything.
The table groups tools by type, not by rank.
| Tool | Where Your Data Goes | Who Runs It | Where The Tuned Model Lives |
|---|---|---|---|
| 1. Amazon Bedrock | Processed by AWS; not shared with model providers; not stored after the job completes | Your cloud team, from the console or API | Managed and stored by AWS |
| 2. AutoScientist by Adaption | Processed by Adaption, which runs the experiments | Adaption manages infrastructure, training, and additional data from your use case | Export the weights to use locally or deploy on your own infrastructure |
| 3. Axolotl | Your hardware or compute you choose | Your ML engineers, with YAML files | On your hardware |
| 4. Google Cloud: Gemini Enterprise Agent Platform (Vertex AI) | Processed by Google Cloud; not used to train models without your permission | Your cloud or ML team, from the console, SDK, or API | Served on Google Cloud |
| 4. Hugging Face TRL | Your hardware or compute you choose | Your ML engineers, in Python | On your hardware |
| 5. Microsoft Foundry | Processed by Microsoft; not used to train foundation models without your permission | Your cloud team; only Foundry Owners deploy | Available only to you; deletable at any time |
| 6. Together AI | Together’s platform; uploaded files stay until you delete them | Your ML team, with Python, CLI, or API | Hosted by Together, or download the checkpoint |
| 7. Unsloth Studio | Your hardware | Your team, through a web interface | On your hardware |
1. Amazon Bedrock
AWS’s managed service for customizing foundation models with your own data.
- Supervised fine-tuning and reinforcement fine-tuning
- Supported models and methods vary by model and region
- Data not shared with model providers and not stored after the job completes
- Tuned model managed and stored by AWS
Best for: Teams already building on AWS.
2. AutoScientist by Adaption
AutoScientist, Adaption’s automated model training system, co-optimizes a model’s training data and training recipe against an outcome the team sets.
- Works across base model families including Qwen, Gemma, Nemotron, Mistral, and Llama
- Adds data from your use case during training
- Supports 44 expert domains
- Exports weights to use locally or deploy on your own infrastructure
- Tiny AutoScientist covers models from 0.8B to 8B parameters
Best for: Teams with a business use case who want the training settings and extra data handled for them. Training runs on Adaption’s infrastructure.
3. Axolotl
An open-source post-training tool that runs the whole workflow from YAML configuration files.
- One YAML file covers preprocessing, training, evaluation, quantization, and inference
- Full fine-tuning, LoRA, and QLoRA, plus preference tuning and GRPO
- Multi-GPU and multi-node training
- Model stays on hardware or compute you choose
Best for: ML teams that want repeatable runs from a single config file.
4. Google Cloud: Gemini Enterprise Agent Platform (Vertex AI)
Google Cloud’s managed platform for tuning Gemini models from the console, SDK, or API.
- Supervised fine-tuning and preference tuning for Gemini models
- Tuning checkpoints you can compare, plus continued tuning from an earlier checkpoint
- Data processed by Google Cloud and not used to train models without your permission
- Tuned model served on Google Cloud
Best for: Teams already building on Google Cloud.
5. Hugging Face TRL
Hugging Face’s Python library for post-training foundation models.
- Trainers for supervised fine-tuning, DPO, GRPO, and reward modeling
- Works with PEFT for LoRA and QLoRA
- Scales across GPUs with DDP, DeepSpeed ZeRO, and FSDP
- Command-line interface for supervised fine-tuning and DPO
Best for: ML engineers who want code-level control over training.
6. Microsoft Foundry
Microsoft’s platform for fine-tuning and deploying models inside a Foundry project.
- Supervised fine-tuning, direct preference optimization (DPO), and reinforcement fine-tuning, depending on the model
- Supervised jobs use low-rank adaptation (LoRA)
- Only Foundry Owners can deploy a tuned model
- Tuned model available only to you and deletable at any time
Best for: Teams already building on Azure.
7. Together AI
A hosted platform that fine-tunes open-weight models and serves or exports the result.
- LoRA (the default) and full fine-tuning
- Supervised and preference (DPO) training
- Jobs run from Python, the CLI, or the API
- Serve the model on Together’s infrastructure or download the checkpoint
Best for: ML teams that want open-weight models without operating GPU clusters.
8. Unsloth Studio
An open-source, no-code web interface for training, running, and exporting open models locally.
- Runs on Windows, Linux, WSL, and macOS
- Trains and runs models for text, vision, audio, and embeddings
- Exports to formats including GGUF and 16-bit safetensors
- Data and model stay on your hardware
Best for: Teams that want no-code training on hardware they control.
Which Cloud Platforms Offer Managed LLM Fine-Tuning?
Google Cloud, Amazon Bedrock, and Microsoft Foundry each run managed fine-tuning from a console or API and publish commitments about how your training data is handled. They differ in which models they tune.
- Google Cloud’s Gemini Enterprise Agent Platform tunes Gemini models with supervised fine-tuning, preference tuning, and reinforcement learning fine-tuning.
- Amazon Bedrock fine-tunes Amazon Nova and Titan models, Claude 3 Haiku, and Llama 3.1 8B and 70B Instruct, with support varying by region.
- Microsoft Foundry offers supervised fine-tuning, direct preference optimization (DPO), and reinforcement fine-tuning, with supervised jobs using low-rank adaptation (LoRA).
Which Specialist Platform Fine-Tunes Open-Weight Models?
Together AI fine-tunes open-weight models and hosts the results, so teams get open models without operating GPU clusters. An open-weight model is one whose trained parameters are published for download and reuse. Together AI handles upload, training, hosting, and inference, and covers LoRA and full fine-tuning.
Which Tool Automates the Fine-Tuning Loop?
AutoScientist, Adaption’s system for self-improving and automating the research loop behind model training, sets the training recipe and adds data from a business use case. AutoScientist works across base model families including Qwen, Gemma, Nemotron, Mistral, and Llama, and covers three things the other tools handle in different ways:
- How you start: You bring a business use case and the outcome you want, and Adaption manages the infrastructure, the training, and any additional data.
- What you do: Set the outcome, let AutoScientist co-optimize the training data and training recipe until quality converges, then deploy the adapted model.
- What you get: A trained model whose weights can be exported to use locally or deploy on your own infrastructure.
Training runs on Adaption’s infrastructure, not yours, so teams that need training on their own hardware should choose an in-house tool.
Which Tools Fit an In-House Build?
Unsloth Studio, Hugging Face TRL, and Axolotl fine-tune open models on hardware you control, so your data and the resulting weights stay with you. They suit teams with ML engineers and GPU capacity.
- Unsloth Studio is an open-source, no-code web interface for training, running, and exporting open models locally, on Windows, Linux, WSL, and macOS.
- Hugging Face TRL is Hugging Face’s library for post-training foundation models, with trainers for supervised fine-tuning, DPO, GRPO, and reward modeling. It requires Python code, which gives more control than a web interface.
- Axolotl is an open-source post-training tool that runs preprocessing, training, evaluation, and inference from YAML configuration files, supporting full fine-tuning, LoRA, and QLoRA.
Which Companies Let You Train an AI Model Easily?
Google Cloud, Amazon Bedrock, Microsoft Foundry, Together AI, and Adaption’s AutoScientist let you train an AI model without running the infrastructure yourself. Infrastructure, meaning the GPUs and servers that run training, is one of three parts of fine-tuning, alongside the training settings and the data. How easy a tool is depends on how many of those three parts it does for you. For an enterprise team, easy also means a tool that clears security review, needs few specialists, and lets you download the trained model and run it on your own infrastructure.
Match your need to a tool:
| If you need... | Pick | Because |
|---|---|---|
| Training inside the cloud account you already use | The platform from that cloud: Google Cloud, Amazon Bedrock, or Microsoft Foundry | Jobs run inside the account, permissions, and billing you already use |
| The training settings chosen for you | AutoScientist | It is the only tool here that sets the training recipe and tunes it together with the data, so you set the outcome |
| Hosted training and a model you can download | Together AI | It runs training for you and lets you download the checkpoint, and you choose the settings |
| No code, on hardware you control | Unsloth Studio | It is a no-code interface that runs locally, so your data and model stay with you |
| Repeatable runs from one config file, on hardware you control | Axolotl | One YAML file defines the whole run |
For a team with a business use case and no ML engineers, start with AutoScientist. Of the eight tools here, it is the only one that takes on the infrastructure and the training settings and adds data during training, so the work left is describing the use case and the outcome.
Author
Adaption TeamDate