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

Together AI expands fine-tuning service with more models, live metrics, and finer controls Webflow Analyze/Optimize tracking bridge -->

🚀 DeepSeek V4 Pro 0813 vs. GPT-5.6 Sol on DeepSWE →

📈 GLM-5.3 vs. GLM 5.3 Flash on DeepSWE →

⚡ On-demand B200s now available on Together GPU Clusters →

🚀 Now serving MiniMax-M3 for efficient inference →

All blog posts

Fine-Tuning

Published 9/11/2026

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

Support for the latest open-weight models, live experiment tracking, finer controls over training and data processing, and lower prices on selected models.

Authors

Artem Chumachenko, Egor Timofeev, Jasmine Li, Ruslan Khaidurov, Nikita Smetanin, Sergei Vorobyov, Arseniy Belorukov, Denis Fedorenko, Alex Moldovan, Shadi Mokhtar, Gleb Vazhenin, Sonny Khan, Adee Feiner, Jen Wu, Max Ryabinin

Table of contents

40+ Models Chosen for Production...40+ Models Chosen for Production...40+ Models Chosen for Production...

Turning an open-weight model into a high-performing model for your task takes a sequence of well-measured experiments. Teams need to understand what the model will train on, follow how each run is progressing, adjust the training recipe, and identify which checkpoint performs best. Today, we're expanding Together Fine-Tuning across that workflow. This release adds support for the latest open-weight models, live experiment tracking, finer controls over training and data processing, and lower prices on selected models. You can also inspect and validate a dataset before training, compare runs while they are in flight, and stop when validation loss plateaus. New models New open-weight models are arriving at an unprecedented pace, with nearly every release raising the bar for quality, efficiency, or both. Teams want to bring these advances into production quickly, without spending weeks on the technical challenges of each new architecture. For tasks that need frontier performance, you can turn to the largest and strongest models such as GLM-5.3 and Kimi K2.7. For example, GLM-5.3 scores 88.2 on Terminal-Bench 2.1, within one point of the leading proprietary models. Other teams prioritize performance and costs, choosing models such as Qwen 3.8-27B or Gemma 4. For local and on-device applications, the Qwen 3.5 family offers options ranging from 0.8B to 9B parameters. Together Fine-Tuning supports all these models across a wide range of scales, with leading optimizations across systems and algorithms built in. Submit a training job, and we handle the architecture-specific complexity behind the scenes. The latest additions include: 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 Gemma 4-26B-A4B

View the full list of supported models along with their context lengths in our documentation. Experiment tracking The sooner a run's trajectory is visible, the sooner the next decision can be made - adjust the data, change the hyperparameters, or let it finish. Fine-tuning jobs now record metrics at every training and evaluation step and expose them directly through the Together API, CLI, and UI dashboard. Metrics, job status, and artifacts stay in one place, so you can remain on the same platform for the entire journey of model development. Every run captures the core training signals: loss, gradient norm, learning rate. When you add a validation set, the evaluation loss and metrics will be tracked alongside training metrics. In the dashboard, every job gets a Metrics tab with curves that update while the run is still in flight - and you can select several jobs and chart them together on one set of axes, so a sweep reads as a single picture instead of a dozen browser tabs. You can also do the same via our Python SDK: one command, curve in the terminal, no context switch.

The API returns the underlying series - every recorded step, filtered by range or downsampled. This way, you can pull metrics into a notebook, your internal dashboards you already have, or any other destination. See more details in our documentation. More controls over training Expert LoRA In a Mixture-of-Experts model, over 90% of the parameters, and most of what the model actually knows , live in the expert layers. Standard LoRA fine-tuning never touches them: adapters attach to attention while the experts stay frozen. You can now put LoRA adapters on the experts themselves, so your job can train the layers where the knowledge lives. The difference shows up wherever a task needs the model to learn something genuinely new. We taught models 200 invented facts (so the base models know none of them): adapters that include the expert layers recalled up to 89% of the new knowledge, while attention-only adapters topped out at 15% on the same model. They also kept more of the model's existing knowledge and won MMLU-Pro outright, 75.3% to 71.5%.

The training dynamics explain the gap: with attention-only adapters, a growing share of routed experts simply falls out of use during fine-tuning, while expert adapters keep the whole mixture engaged. Enabling expert adapters is simple: just add the expert modules to lora_trainable_modules . The rest of the fine-tuning workflow remains the same as it does for any other LoRA job . Early stopping Fine-tuning jobs can now automatically finish when they stop improving. That means you get the best model, not just the last one, and you don't pay for training that no longer helps. Enable early stopping, provide a validation set, and we'll watch your validation loss at every evaluation. When the curve plateaus, we halt the run, keep the checkpoint with the best validation loss as your final model, and automatically refund every training step you didn't use. Patience, sensitivity to small improvements, and warmup are all tunable when you want finer control. The full parameter reference is in the docs . Arbitrary batch size support Some jobs need batch sizes that can't fit in GPU memory: long sequences, large models or a recipe tuned for a big global batch. You can now train with whatever effective batch size you want regardless of what fits in a single pass. Set gradient_accumulation_steps and gradients accumulate across micro-batches before each optimizer update, giving an effective batch of batch_size x gradient_accumulation_steps. Details in the docs ....