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

Together AI has rolled out a significant expansion of its fine-tuning platform, adding support for a wider set of open-weight models alongside new tools for monitoring and controlling training jobs.

What's new

According to Together's announcement, the service now offers "support for the latest open-weight models," including GLM-5.3, DeepSeek-V4-Flash variants, Kimi K2.7-Code, and multiple Qwen 3.5 and 3.8 parameter sizes. On visibility, the company says "fine-tuning jobs now record metrics at every training and evaluation step and expose them directly through the Together API, CLI, and UI dashboard," replacing what had previously been a more opaque training process for customers.

The release also introduces new training controls: "Expert LoRA" for mixture-of-experts models, an early-stopping mechanism that halts training when validation loss plateaus and issues automatic refunds for the unused compute, and gradient accumulation support for arbitrary batch sizes. On the data side, users can now "inspect the tokenized data before starting a job," with previews showing the exact tokens and labels a model will train on before any compute is spent.

Together is also cutting prices on fine-tuning: the announcement cites reductions of 30% to 70% across selected models, with Qwen3.5-9B LoRA SFT dropping from $0.48 to $0.34 per million tokens as one specific example.

Context

The update lands one day after Together AI announced preemptible GPU compute at half the on-demand price, part of a broader push by the company to compete on both raw compute economics and developer tooling against rivals like Fireworks and Baseten in the open-model hosting and fine-tuning market. As open-weight model families from Alibaba's Qwen, Zhipu's GLM, Moonshot's Kimi, and DeepSeek continue to multiply and iterate quickly, demand has grown for platforms that can keep pace with day-one support and give teams enough visibility into training runs to justify the spend.

Why it matters

Fine-tuning tooling is a quieter but consequential layer of the AI stack: better visibility and controls reduce the wasted-compute cost of failed or over-long training runs, and mixture-of-experts-specific tooling like Expert LoRA matters more as MoE architectures become the default for new open-weight releases. Combined with the steep price cuts, the update makes Together a more credible option for teams customizing open models rather than defaulting to closed-model APIs, reinforcing the broader trend of open-weight fine-tuning becoming cheaper and more observable rather than remaining a black box.