Baseten puts Loops in public preview: Fine-tune open AI models without managing servers

AI company Baseten offers a tool called Loops that makes it possible to fine-tune open language models without having to run the heavy technical infrastructure yourself. According to a report on Baseten's launches, Loops was already in public preview, an open trial period, and was scheduled for general availability on Friday, November 6, 2026.

What is Loops?

Loops is a Python SDK, a toolkit for programmers, for so-called LoRA fine-tuning. LoRA is a method for adapting an existing AI model to a specific task by training a small extra layer instead of retraining the entire model from scratch.

Developers write their own training scripts for two types of training: supervised fine-tuning (SFT), where the model learns from examples, and reinforcement learning (RL), where the model is rewarded for good answers. Baseten, in turn, handles the trainer, the sampler and the rest of the infrastructure.

The SDK has four main operations: computing loss and accumulating gradients, updating the adapter, generating text from the current model, and saving the training state so it can be resumed later.

According to the documentation, the API is compatible with Tinker, so a script that imports tinker can run on Loops without rewriting the calling code. Loops supports LoRA training of open models from Qwen, Google and DeepSeek, among others, including models that understand both images and text.

Two modes: serverless and dedicated

In serverless mode, you share capacity with others. Billing is per token at model-specific rates for training and sampling, and you don't need to reserve a GPU. Idle runs cost nothing, and time between operations, for example when reviewing results or adjusting a reward function, is not billed. The documentation does not list specific prices.

The shared capacity can fill up. When that happens, the system returns a 429 error and retries up to ten times. Users can ask for more concurrent runs or larger models through Baseten's support channels.

Serverless only works for models marked "Serverless and dedicated," and with certain sequence lengths. If a configuration isn't supported, it automatically falls back to dedicated mode, or it fails if only serverless is required.

Dedicated mode provides full model support and is billed by GPU time. Here you control replicas and capacity.

According to a report on Baseten's product page, models trained with Loops can also be moved directly into Baseten Dedicated Inference with a single command.