Together AI fine-tuning adds expert LoRA, live metrics, early stopping, and price cuts

Together AI expanded its fine-tuning service on September 11, 2026, adding support for the latest open-weight models, live experiment tracking, finer controls over training and data processing, and lower prices on selected models.

New models and experiment tracking

The service now supports models including GLM-5.3, GLM-5.2, GLM-5.1, DeepSeek-V4-Flash-0731, DeepSeek-V4-Flash, Kimi K2.7-Code, Kimi K2.6, Qwen 3.8-27B, Qwen 3.6-35B-A3B, Qwen 3.6-27B, Qwen 3.5-27B, Qwen 3.5-9B, Qwen 3.5-4B, Qwen 3.5-2B, Qwen 3.5-0.8B, Gemma 4-31B, and Gemma 4-26B-A4B. Together AI said GLM-5.3 scores 88.2 on Terminal-Bench 2.1, within one point of leading proprietary models. The Qwen 3.5 family covers 0.8B to 9B parameters for local and on-device applications.

Fine-tuning jobs now record metrics at every training and evaluation step, exposed through the Together API, CLI, and UI dashboard. Each run captures loss, gradient norm, and learning rate. When a validation set is added, evaluation loss and metrics are tracked alongside training metrics. The dashboard shows a Metrics tab with curves that update while the run is in flight, and multiple jobs can be charted together on one set of axes. The API returns the underlying series, filtered by range or downsampled, so metrics can be pulled into notebooks or internal dashboards.

Expert LoRA and early stopping

In a Mixture-of-Experts model, over 90% of parameters live in the expert layers, and standard LoRA fine-tuning leaves them frozen. Together AI now allows LoRA adapters on the expert layers themselves. The company reported that when models were taught 200 invented facts, adapters including expert layers recalled up to 89% of the new knowledge, while attention-only adapters topped out at 15% on the same model. Expert adapters also kept more existing knowledge and won MMLU-Pro outright, 75.3% to 71.5%. Enabling expert adapters requires adding the expert modules to the lora_trainable_modules option.

Early stopping lets jobs finish automatically when validation loss plateaus. The run halts, the checkpoint with the best validation loss becomes the final model, and Together AI refunds every training step not used. Patience, sensitivity to small improvements, and warmup are tunable. The service also supports arbitrary batch sizes through gradient accumulation, letting users set an effective batch size that exceeds what fits in a single pass.

Price reductions and data controls

Together AI dropped training prices for most supported models, with savings starting at 30% and reaching 70% for the gpt-oss series. Per 1 million tokens for LoRA adapters, Qwen 3.5-9B SFT dropped from 0.48 to 0.34, Qwen 3.8-27B SFT from 1.50 to 1.05, Gemma 4-31B SFT from 1.50 to 1.05, gpt-oss-120b SFT from 5.00 to 2.50, and gpt-oss-20b SFT from 1.50 to 0.40. DPO prices dropped proportionally.

Data processing now includes pre-flight file validation that runs as soon as a JSONL file finishes uploading, checking message structure, tool definitions, reasoning fields, multimodal content, preference-pair structure, and sample weight types. Users can inspect tokenized data before starting a job, seeing token IDs, token strings, per-token labels, and truncation status. Every job exposes a subset of its fully prepared dataset through the Python SDK or a download link. Per-example weighting is supported by adding a weight field to any JSONL example. Packing controls remain available, though packing is enabled by default for throughput.

Together AI also shared a case study on Adaption, a research company using the platform to automate training loops for open models up to 1 trillion parameters, with evaluation run on Together's dedicated endpoints.