DeepSeek V4 Flash 0731 opens for fine-tuning on Together AI

The model DeepSeek V4 Flash 0731 is now available for fine-tuning on the Together AI platform, allowing developers to adapt it for code, tool use and specialized domains via SFT or DPO, and then run the tuned version on Together’s inference infrastructure in production.

two training methods with clear trade-offs

Together offers two paths: LoRA, the default, trains a small set of adapter weights on top of a frozen base model; it is faster, cheaper and fits most use cases. The second option is full fine-tuning, which updates every weight in the base model, requires more compute resources, but may surpass LoRA when a substantial change in core behavior is required.

data methods and preparation

Data must be provided as a JSONL or Parquet file that matches the task schema; Together’s guide specifies validation rules and examples. The default method is SFT (supervised fine-tuning), which learns from demonstrations with a single target completion per example. Preference fine-tuning (DPO) is also available, aligning the model with rankings of preferred versus non-preferred responses.

advanced capabilities: vision, function calling, reasoning

The advanced guides cover three separate directions: fine-tuning vision-language models on image-text samples; training for reliable function calling; and training a reasoning model with chain-of-thought data. When using LoRA, the rank and targeted modules can also be specified.

monitoring, deployment and local download

During training you can monitor task status, retrieve loss and evaluation metrics at each step. After training, the tuned model is served at a dedicated endpoint on Together, or can be downloaded for local inference. Pricing is based on training tokens and dedicated hosting; full details are on the platform’s pricing page.

external models from the Hugging Face Hub

Beyond the internal catalog, Together allows importing a model from the Hugging Face Hub that is not listed in the official catalog and fine-tuning it on the same infrastructure. The list of supported base models includes context lengths and batch sizes for each model.