Together Fine-Tuning

What it does

Together Fine-Tuning takes a training dataset (JSONL or Parquet, optionally with images for vision-language models, tool-call definitions, or preference pairs) and an open-source base model, and produces a fine-tuned model that can be hosted on a dedicated endpoint or downloaded. It supports supervised fine-tuning and DPO preference training, using either LoRA adapters or full-weight training.

Teams launch jobs from the web console, the API/SDK, or the tg command-line tool, upload data, monitor per-step loss and evaluation metrics, and deploy the result. Its distinguishing points are managed multi-node training for very large models, including 100B+ parameter ones, a bring-your-own model route for Hugging Face Hub models, cost estimates before a job runs, and training on existing agent logs for function calling.

It suits developers and ML teams that want a customized open model without running training infrastructure. Access requires a Together AI account and API key. Billing is per training token with a per-model minimum charge. Only models in the supported list can be fine-tuned, and some cannot be hosted on dedicated endpoints, so check deployability before training. Hosting is billed separately.

What it can do

• LoRA and full fine-tuning of open-source models
• Supervised fine-tuning and DPO preference training
• Vision-language fine-tuning on image and text data
• Tool-calling and reasoning fine-tuning
• Cost estimation before launching a job
• Deployment to a dedicated endpoint or model download

Works with

• Hugging Face Hub

Example workflows

Domain-tune a model with LoRA

Input: a train.jsonl file of task examples. Steps: upload it with tg files upload, create a LoRA job on a chosen base model, check progress with tg fine-tuning retrieve. Output: a fine-tuned model ready to deploy on a dedicated endpoint.

Train tool calling from agent logs

Input: existing agent logs containing function definitions and tool calls. Steps: format them as a fine-tuning dataset, estimate cost in the UI or CLI, launch the job. Output: a model trained for more precise tool execution.

Before you choose

• Cost is per training token (dataset size times epochs plus evaluation tokens), with a minimum charge per model and job.
• Some fine-tunable models cannot be deployed on dedicated endpoints; verify before training.
• Dedicated hosting is a separate cost from training.

Listed under

• Coding
• Developers
• Dpo
• Enterprise AI
• Fine-tuning
• LLM training
• Open-source models

Get started

Requires a Together AI account; upload data, pick a model, launch.

1. Create a Together AI account and obtain an API key.
2. Prepare and upload a JSONL or Parquet training file following the data preparation guide.
3. Check the supported-models list and cost estimate, then launch a LoRA or full job.

Billed per 1M training tokens (dataset size x epochs plus evaluation tokens), with a per-model minimum charge per job. See the pricing page rate table for current rates.

Source: together.ai/pricing, checked 29 Oct 2026. Prices can change; check before you buy.

Sources for this overview

Product pages, documentation, and pricing checked 29 Oct 2026. Each source supports part of the overview; availability and prices can change.

• Fine-Tuning | Together AI ↗www.together.ai
• Pricing | Together AI ↗www.together.ai
• Overview - Together AI docs ↗docs.together.ai
• Supported models - Together AI docs ↗docs.together.ai

Company website

together.ai: product, pricing and policy pages

YC directory mirror

Batch and company name

Social accounts

Accounts the company links, and its posts on X

Human review

Checked by a person

knowledge_base

Last checked 29 Oct 2026. Every field keeps its source and date; unknown means we couldn’t confirm it.