Together Fine-Tuning
Together AI's managed fine-tuning for open models: LoRA or full, SFT and DPO, vision and tool-calling data, billed per token
Overview
Together Fine-Tuning is the managed fine-tuning service of Together AI, the open-model cloud. You upload a dataset, pick an open model and Together runs the job with no GPUs to manage: LoRA (the default) or full fine-tuning, supervised fine-tuning (SFT) or preference tuning with DPO, plus vision data, function-calling data and reasoning traces. It accepts open models from the Hugging Face Hub, including very large mixture-of-experts models, and when the job is done you can serve the model on a Together endpoint or download the weights to run anywhere. Billing is per token processed (training tokens times epochs, plus evaluation tokens) at per-model rates, with a $4 minimum per job on most models; failed jobs are refunded. For custom RL and SFT loops on dedicated GPUs, Together launched a separate Together Post-Training product in October 2026. Together AI was founded in 2022 in San Francisco and raised a $305M Series B at a $3.3B valuation in February 2025. Also from Together AI: Together AI inference.
The Verdict
Who Should Use Together Fine-Tuning?
Best For
- Teams that want to tune open models without running GPUs
- Projects that need to own and download the tuned weights
- Apps already serving models on Together AI
Not Ideal For
- Tuning closed models such as GPT or Claude
- Fully custom RL algorithms (use Together Post-Training or open-source tools)
- Beginners: setup expects technical skill
What's Great
- LoRA and full fine-tuning, SFT and DPO, vision and tool-calling data
- Tuned weights can be downloaded, not locked in
- Clear per-token pricing with refunds for failed jobs
- Customer XY.AI Labs reports 2 to 3x cost savings (vendor case study)
Pricing
View all features & details
Key Features
- LoRA (default) and full fine-tuning
- SFT and DPO preference tuning
- Vision (image plus text), function-calling and reasoning data
- Open models from the Hugging Face Hub, incl. large MoE models
- Deploy to a Together endpoint or download the weights
- Data residency options in North America, Europe and Asia / Middle East
Billing
- Tokens = epochs x training tokens + evaluations x validation tokens
- Cancelled jobs: pay only for completed steps
- Failed jobs: fully refunded
- Hosting billed separately per minute on dedicated endpoints
Real-World Usage
Customer Stories
- XY.AI Labs: 2 to 3x cost savings, accuracy up from 77% to 87%
- Slingshot AI and Protege AI also featured
Company
- $305M Series B at $3.3B, led by General Catalyst and Prosperity7
- Together Post-Training launched Oct 2026
How It Compares
| Feature | Together Fine-Tuning | Axolotl | Unsloth |
|---|---|---|---|
| Type | Managed service | Open source, self-run | Open source, self-run |
| GPUs | Managed by Together | Your own or rented | Your own or rented |
| Methods | LoRA and full; SFT, DPO | LoRA to full; DPO, KTO, ORPO, GRPO | LoRA, QLoRA, full; DPO, GRPO |
| Pricing | Per 1M tokens, per model | Free | Free core; paid tiers by quote |
Sources: Together Fine-Tuning · Pricing · Docs · Pricing docs
Is this your tool? Add a badge