Together AI expands fine-tuning service with more models, live metrics, and finer controls

Together AI announced an expansion of its Together Fine-Tuning service, adding support for new open-weight models along with live experiment tracking, finer training controls, and lower prices on selected models.

Newly supported models include GLM 5.3, GLM-5.2, GLM-5.1, DeepSeek-V4-Flash-0731, DeepSeek-V4-Flash, Kimi K2.7-Code, Kimi K2.6, several Qwen 3.8/3.6/3.5 variants ranging from 0.8B to 27B parameters, and Gemma 4-31B and Gemma 4-26B-A4B. Together AI states GLM-5.3 scores 88.2 on Terminal-Bench 2.1, within one point of leading proprietary models.

Fine-tuning jobs now record metrics at every training and evaluation step, exposed through the Together API, CLI, and UI dashboard, capturing loss, gradient norm, and learning rate, plus evaluation loss and metrics when a validation set is provided.

A new Expert LoRA option places LoRA adapters on the expert layers of Mixture-of-Experts models. Together AI reports that on a test of 200 invented facts, expert-layer adapters recalled up to 89% of the new knowledge versus 15% for attention-only adapters, and scored 75.3% versus 71.5% on MMLU-Pro.

Early stopping halts runs when validation loss plateaus, keeping the best checkpoint and refunding unused training steps. Arbitrary effective batch sizes are supported via gradient_accumulation_steps. Users can preview tokenized data before training, add per-example sample weights in JSONL files, run pre-flight server-side file validation on upload, and control sequence packing.

Training prices for LoRA adapters per 1 million tokens were reduced, with savings from 30% up to 70% for models such as the gpt-oss series. For example, Qwen3.5-9B SFT drops from 0.48 to 0.34, and gpt-oss-120b SFT from 5.00 to 2.50.

Together AI also described a case study with Adaption, whose AutoScientist trains and evaluates models up to 1T parameters using Together's fine-tuning platform and dedicated endpoints. Adaption co-founder Sara Hooker said the partnership allows the company to provide self-improving intelligence at scale.

Together AI said that intermediate LoRA adapters will become deployable while fine-tuning is still in progress, with initial support planned for GLM-5.3, followed by Kimi K3.