TogetherFine-Tuning

Train, improve & deploy high-quality, fast models that excel in your specific domain.

Custom models that are faster, more accurate, cheaper, and 100% yours

Together Fine-Tuning allows you to train open-source models with your data to create models that excel at specific tasks, match your tone of voice & more.

Task-specific models
Improve overall quality by fine-tuning a model to provide high-quality results customized for specific tasks and domains.
Smaller & faster at lower cost
Create smaller fine-tuned models that match the quality of large models with much faster performance and lower cost.
Deploy & download
Once your model is created, seamlessly deploy and run inference on Together Cloud or download the resulting checkpoint.

"After thoroughly evaluating multiple LLM infrastructure providers, we’re thrilled to be partnering with Together AI for fine-tuning. The new ability to resume from a checkpoint combined with LoRA serving has enabled our customers to deeply tune our foundation model, ShieldLlama, for their enterprise’s precise risk posture. The level of accuracy would never be possible with vanilla open source or prompt engineering."

‍

- Alex Chung, Founder of Protege AI

Train models that adapt & evolve with your users

Fine-tune leading open-source models to capture your domain expertise and continuously adapt them as your app evolves.

Run LoRA or full fine-tuning jobs

Start full fine-tuning jobs to create new custom models trained on your data.

For a faster, memory-efficient approach, use LoRA fine-tuning, which produces small adapters you can use when running inference.


together.fine-tuning.create(
    "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B", # Base model
    "training-file.jsonl", # Dataset
    training_method="lora", # Optionally set LoRA training method
)

Meet user preferences with Direct Preference Optimization

Align models with human preferences using preferred vs. non-preferred responses to fine-tune output quality.

By training directly on preferences, DPO teaches your model to distinguish good responses from bad ones, yielding more helpful, human-centric outputs.


together.fine-tuning.create(
    "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B", # Base model
    "preference_dataset.jsonl", # Dataset with preferred/rejected messages
    training_method="dpo", # Set DPO as the training method
    dpo_beta=0.2 # Set DPO beta
)

Evolve over time with Continued Fine-Tuning

Adapt an already fine-tuned LLM to new tasks, languages or data without losing the skills it has already learned.

Building on an existing model’s knowledge saves time and compute resources versus training from scratch.

Continued fine-tuning guards against catastrophic forgetting, preserving prior strengths while adding new skills.

# Fine-tune from the base model
together.fine-tuning.create("deepseek-ai/DeepSeek-R1-Distill-Llama-70B", "adapt_dataset.jsonl")

# Get checkpoint ID
{"id": "ft-0451..."}

# Fine-tune from checkpoint
together.fine-tuning.create(from_checkpoint="ft-0451...", "finetune_dataset.jsonl")

Read our Docs

Deploy seamlessly to Together Cloud

01
Serverless LoRA
Run LoRA inference on supported base models directly on Together serverless endpoints to get maximum flexibility.
Read our Docs
02
Dedicated Endpoints
Deploy your LoRA or full fine-tuning models on Dedicated Endpoints for full control & better price-performance at scale.
Read our Docs

Get started

Read our quickstart guides to start fine-tuning models with your own data today.

Fine-Tuning Quickstart Guide

TOGETHER AI

Direct Preference Optimization Guide

Ivan Provilkov, Zain Hasan, Max Ryabinin

Continued Fine-tuning of LLMs

Artem Chumachenko, Zain Hasan, Max Ryabinin

Start Training Your Custom Models Today!

Start fine-tuning

TogetherFine-Tuning

Custom models that are faster, more accurate, cheaper, and 100% yours

Task-specific models

Smaller & faster at lower cost

Deploy & download