Together AI expanded its fine-tuning platform on September 11, 2026, rolling out support for 17 new open-weight models, live experiment tracking inside its dashboard and API, a technique for training the expert layers of Mixture-of-Experts models, and training price cuts reaching 70% on select models. The changes, detailed in a company blog post published the same day, touch nearly every stage of the fine-tuning workflow, from inspecting a dataset before a job starts to deciding when a run should stop.
The release folds seven separate additions into one push: new model support, a live metrics dashboard, an MoE-specific LoRA variant called Expert LoRA, automatic early stopping, arbitrary batch sizes, dataset preview and validation tools, and the price reductions. Together also published a case study describing how AI research startup Adaption uses the platform to run an automated training loop.
This is same-day coverage. No independent outlet had reported on the update as of publication, so everything below traces back to Together's own blog post and documentation, with the company's benchmark claims flagged as such rather than treated as confirmed fact.
Quick Take
- Together AI added 17 open-weight models to its fine-tuning platform, including GLM 5.3, DeepSeek-V4-Flash, Kimi K2.7-Code, and the full Qwen 3.5 through 3.8 lineup.
- Training prices dropped 30% to 70% on select models. The steepest cut in Together's published table takes gpt-oss-20b from $1.50 to $0.40 per million tokens for LoRA SFT.
- A new technique called Expert LoRA trains the expert layers inside Mixture-of-Experts models. Together's own internal test showed it recalling up to 89% of injected knowledge versus 15% for standard adapters, a number the company has not had independently verified.
More Models, Faster Onboarding
Together added 17 open-weight models to its fine-tuning lineup, spanning frontier-scale options like GLM 5.3 and Kimi K2.7-Code down to small, on-device sizes in the Qwen 3.5 family, which runs from 0.8B to 9B parameters. Together points to GLM-5.3's score of 88.2 on Terminal-Bench 2.1, which the company says lands within a point of leading proprietary models. That benchmark score comes from GLM's maker, Z.ai, not from Together's own testing.
Other additions include DeepSeek-V4-Flash, DeepSeek-V4-Flash-0731, Kimi K2.6, three sizes of Qwen 3.6, Qwen 3.8-27B, and two Gemma 4 variants. Together's supported models documentation lists the full set along with context lengths.
Live Metrics Instead of Blind Waiting
Every fine-tuning job now streams loss, gradient norm, and learning rate at each training and evaluation step, viewable in a dashboard Metrics tab, through the Together API, or via the Python SDK. Multiple jobs can be charted on the same axes, so a hyperparameter sweep shows up as one comparison instead of a dozen open browser tabs.
Finer Controls Over Training
Expert LoRA Targets the Part of the Model That Actually Learns
Standard LoRA fine-tuning attaches adapters to a model's attention layers and leaves the expert layers, where Together says most of a Mixture-of-Experts model's parameters and knowledge live, untouched. The new Expert LoRA option lets adapters train those expert layers directly.
In an internal test, Together taught models 200 invented facts the base models could not have already known. Adapters that included the expert layers recalled up to 89% of the new information, compared with 15% for attention-only adapters on the same model. The expert-layer adapters also scored higher on MMLU-Pro, 75.3% versus 71.5%. Those figures come entirely from Together's own benchmarking. Together's documentation on targeting MoE expert layers covers setup.
Early Stopping and Flexible Batch Sizes
Jobs can now stop automatically once validation loss plateaus. Together keeps the checkpoint that scored best on validation and refunds unused training steps. A separate change lets developers set an effective batch size larger than what fits in GPU memory at once, using gradient accumulation to combine smaller micro-batches into a single optimizer update.
Cleaner Data Handling
Together added three changes aimed at what happens before a job starts. A new preview tool shows exactly how a dataset gets tokenized and packed, so a misapplied chat template or a truncated row surfaces before training rather than after. Per-example sample weighting lets specific rows in a dataset carry more or less influence on the loss. And server-side file validation now checks a dataset's structure as soon as it finishes uploading, rather than after a job has already queued and started, flagging the index of any malformed example.
Together also made sequence packing, which concatenates short training examples to cut down on wasted computation, fully optional, since packing can leave too few optimization steps per epoch on small datasets.
Prices Fall Across Most Models
Together cut training prices on most supported models, with savings the company puts at 30% to 70%. The steepest cuts land on OpenAI's open-weight gpt-oss models.
Prices shown are for LoRA training per million tokens. The gpt-oss-20b cut works out to roughly 73%, the largest drop in the table Together published. Together's fine-tuning pricing page has the full list.
A Real Customer: Adaption's AutoScientist
Together's post includes a case study on Adaption, a research company building models designed to keep learning from live data instead of relying on static retraining cycles. Its AutoScientist system automates the loop of picking a training recipe, running it, and evaluating the result, using Together's fine-tuning platform for training and Together's dedicated endpoints for evaluation. Adaption says the system supports both LoRA and full fine-tuning on open models up to 1 trillion parameters.
Adaption co-founder Sara Hooker is quoted in the post crediting the partnership with letting the company offer self-improving models at that scale. Worth noting: it's a customer quote supplied for Together's own announcement, not independent verification of the results.
What's Coming Next
Together says a future update will let developers deploy an intermediate LoRA adapter while its training job is still running, instead of waiting for the full run to finish. Initial support is planned for GLM-5.3, followed by Kimi K3. Together has not given a release date.
Why It Matters
This is Together trying to close the gap between an open-weight model landing on Hugging Face and a business actually being able to use it in production. Models like GLM 5.3 and Kimi K2.7 now show up on the fine-tuning platform fast enough to matter, and the price cuts on gpt-oss in particular land right as OpenAI's open-weight releases have been drawing developers looking for a cheaper alternative to its hosted API.
The Expert LoRA numbers are the part I'd wait on before repeating as fact. An 89% knowledge recall rate against a 15% baseline is a big gap, and it lines up with something people doing MoE fine-tuning have complained about for a while: adapters that never touch the expert layers barely move the needle on anything the base model doesn't already know. But it's Together's own test, on a benchmark Together built and ran. I'd want to see someone outside the company reproduce that 200-fact test before treating the number as settled.
What to Watch
Watch for outside developers running their own knowledge-injection tests against Expert LoRA and publishing results that either back up or undercut the 89-versus-15 gap. Watch whether the price cuts extend past the five models Together listed as more open-weight releases land. And watch the GLM-5.3 rollout of deploy-while-training, since that's the feature with the clearest effect on how fast teams can actually iterate.
Key Takeaways
- Together AI added 17 open-weight models to its fine-tuning platform, including GLM 5.3, Kimi K2.7-Code, and the full Qwen 3.5 through 3.8 lineup.
- Training prices fell 30% to 70% on select models. gpt-oss-20b dropped from $1.50 to $0.40 per million tokens for LoRA SFT, the steepest cut Together published.
- Expert LoRA trains a Mixture-of-Experts model's expert layers instead of just attention. Together's own internal test showed 89% knowledge recall versus 15% for standard adapters, a figure not yet independently verified.
- New early stopping, dataset preview, and pre-flight validation tools aim to catch data and training problems before a job wastes compute.
FAQ
What did Together AI actually change in its fine-tuning service?
Together added support for 17 new open-weight models, live experiment tracking in its dashboard and API, a Mixture-of-Experts-specific technique called Expert LoRA, automatic early stopping, dataset preview and validation tools, and training price cuts of 30% to 70% on select models. The changes were detailed in a company blog post published September 11, 2026.
How much cheaper is fine-tuning on Together AI now?
It depends on the model. Together's published table shows LoRA SFT training on gpt-oss-20b dropping from $1.50 to $0.40 per million tokens, while Qwen3.5-9B fell from $0.48 to $0.34. Together says cuts of 30% to 70% apply across most supported models, though it hasn't published a complete list of every affected model.
What is Expert LoRA and how is it different from standard LoRA?
Standard LoRA fine-tuning only adjusts a model's attention layers. Expert LoRA extends that to the expert layers inside Mixture-of-Experts models, which Together says hold most of a model's actual knowledge. In Together's own test, this recalled up to 89% of newly taught information versus 15% for attention-only adapters. That result comes from Together's internal benchmarking and hasn't been independently verified.
Is the deploy-while-training feature available now?
No. Together describes it as coming soon, with initial support planned for GLM-5.3 followed by Kimi K3. The company hasn't given a release date.
