Together AI fine-tuning now accepts any compatible Hugging Face Hub model
Hugging Face
- Developer tools
- Models
- Industry
- #together ai
- #hugging face
- #fine-tuning
Browse Developer tools topic page →
Together AI announced a capability that lets its fine-tuning service pull a compatible model directly from the Hugging Face Hub and train it on Together's infrastructure. Callers specify a catalog base model as a configuration template plus the chosen Hub model, and can optionally upload the result back to a Hub repo. The publishers describe support for CausalLM models under 100B parameters and report early adopter usage.
A Hub model as the training target
Together AI published a post describing a fine-tuning capability that selects the model to train from the Hugging Face Hub. In the shown API call, the model parameter names a model from Together's catalog, while a new from_hf_model parameter names the Hub model. Together explains the catalog model as a template that sets GPU allocation, memory configuration, training pipeline and inference setup, so it should match the custom model in architecture, approximate size and sequence length; the example pairs a SmolLM2 1.7B model with a Llama-family base template. Training data is uploaded as a JSONL file, and a hf_output_repo_name can push the result back to the Hub, with hf_api_token for private repos. The post states that CausalLM models under 100B parameters are generally intended to work.
Scenario: a domain model not in the catalog
A machine learning engineer at a company with a narrow corpus, such as support transcripts from one product line, could try this when the best starting point is a community model that is absent from the vendor's catalog. Rather than standing up training infrastructure, the engineer could pair that Hub model with a same-family catalog template, run a short supervised fine-tuning job and push the checkpoint to a private Hub repo for the team. Iterative improvement is a second plausible use: taking a prior fine-tuned checkpoint, uploading it, and continuing training. Whether this fits depends on the target model matching a template, which Together presents as the condition for optimal results.
How to trial it against a baseline
A team could run a controlled comparison: evaluate the untuned Hub model on a held-out set first to fix a baseline, then run the same training file through the fine-tuning call and re-evaluate with identical prompts and scoring. Start with one template pairing that follows the architecture and size guidance, and log training file size, epochs and the job identifier. Success means the tuned checkpoint beats the baseline on the task metric without regressions on general prompts. Failure checks should include whether the chosen template truly aligns with the custom model and whether data quality, not infrastructure, caps gains. A reviewer should inspect a sample of generated outputs before deployment, since the publisher reports results but no independent comparison is available.
Trade-offs and open questions
The template mechanism is the main trade-off: flexibility extends to Hub models that resemble something in the catalog, while unusual architectures may lack a suitable base and push people back to the intended subset of CausalLM models under 100B parameters. Prerequisites include a JSONL training file, a Hugging Face token for private repositories and a matching output repo name for publishing. Unknowns from this excerpt include supported data formats beyond the example, pricing, limits and failure modes. The post's speed and cost benefits come from Together and named beta users, so treat them as claims to verify. Teams should not use this as a substitute for their own evaluation, and should avoid uploading user data to repositories they do not control.
Context and limitations
- Based on a short publisher excerpt; the linked implementation guide and full documentation were not reviewed.
- Support is described as intended for compatible CausalLM models under 100B parameters, not for every model on the Hub.
- Performance, cost and reliability claims in the post come from the publisher and named beta users, not independent tests.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.