Capability · Framework — fine-tuning
Together Fine-Tuning
Together AI offers a managed fine-tuning product for teams who want to customise open-weight models without owning training infrastructure. You upload a JSONL dataset, pick a base model, and Together runs the job on their GPU cluster, returning either downloadable weights or a hosted inference endpoint. Pricing is per-token-trained and typically 10-100x cheaper than DIY on cloud GPUs for small jobs.
Framework facts
- Category
- fine-tuning
- Language
- API + Python SDK
- License
- commercial
Install
pip install togetherQuickstart
from together import Together
client = Together(api_key='...')
# Upload dataset
file = client.files.upload(file='train.jsonl', purpose='fine-tune')
# Launch job
job = client.fine_tuning.create(
training_file=file.id,
model='meta-llama/Llama-3.1-8B-Instruct',
n_epochs=3,
lora=True
)
print(job.id)Alternatives
- OpenAI fine-tuning — closed-weight models only
- Fireworks fine-tuning service
- Modal / RunPod — DIY with Unsloth or Axolotl
- Predibase — managed open-source fine-tuning
Frequently asked questions
Hosted fine-tuning or DIY?
Hosted wins for one-off jobs, teams without ML-ops, and when you want to deploy immediately as an endpoint. DIY (Unsloth / Axolotl on Modal or RunPod) wins for frequent training, custom algorithms, or when you need to own the weights outright. Together supports both paths.
Can I download the weights afterward?
Yes for LoRA adapters and typically yes for full weights of open models, subject to the base model's license. Always check — some base models (Llama) have acceptable-use restrictions that carry over.
Sources
- Together Fine-Tuning — docs — accessed 2026-04-20