Together AI expands fine-tuning with new models, live tracking, up to 70% price cuts

Together AI has rolled out a major update to its fine-tuning platform, adding support for a slate of new open-weight models, real-time experiment tracking with parallel run comparison and early stopping, and price reductions of 30% to 70% on selected models.

The model lineup is extensive. It includes the GLM 5.1–5.3 series — GLM-5.3 scores 88.2 on Terminal-Bench 2.1, one point shy of the leading proprietary models — two versions of DeepSeek-V4-Flash, Kimi K2.7-Code and Kimi 2.6, the Qwen 3.5 family ranging from 0.8B to 27B parameters alongside Qwen 3.6 and 3.8 in various sizes, and Gemma 4 in two configurations: a 31B dense model and a 26B-A4B mixture-of-experts variant. Together handles the architectural complexity behind the scenes so developers submit a training job and receive a tuned model without wrestling with the implementation details of each architecture.

On the experimentation side, the platform now logs metrics at every training and evaluation step — loss, gradient norm, learning rate — and surfaces them through the API, CLI, dashboard, and Python SDK. Multiple runs can be compared on the same graph while they are still running. Raw series can be pulled into notebooks or internal systems, and runs stop automatically when validation loss plateaus, eliminating the need to monitor dozens of browser tabs.

The most technically significant addition is Expert LoRA. In mixture-of-experts models, more than 90% of parameters and knowledge reside in the expert layers, yet standard LoRA freezes those layers and attaches adapters only to the attention layers. Expert LoRA lets developers attach adapters to the experts themselves, so training reaches the layers where knowledge is actually stored. Together demonstrated the approach on 200 fabricated facts the models had never seen; adapters that included the expert layers remembered them significantly better than the base version.

The 30%–70% price cut on selected models extends the broader trend of falling cloud fine-tuning costs, making large-scale experimentation more accessible to small teams. Combined with dataset preview capabilities before training, the full chain — from data inspection through live tracking to comparison and checkpoint selection — now runs inside a single platform without switching tools.

The cadence of powerful open-weight model releases demands infrastructure that closes the gap between publication and production. Together is positioning itself as the abstraction layer that absorbs shifting architectural complexity — MoE, sparse attention, varying sizes — so teams can focus on data and the training recipe instead of systems engineering.