Muse Spark Open Post-Training: Baseten Loops Takes Checkpoints Straight to Production

Baseten's official tweet announced that Meta's Muse Spark Open can now be post-trained on the Loops SDK. Post-training means continuing to train on pretrained weights using your own data and reward functions. Once a checkpoint is ready, deploy it directly to a production inference endpoint.

Many teams fine-tuning models have to manage training environments and GPU instances themselves, and after training they also need to set up a separate inference service. Loops' approach is for training and sampling to share one session, so you can generate text from the latest weights during training, eliminating the handoff between checkpoint and inference service. The official quickstart documentation provides a complete example; the base model is Qwen, but the workflow applies equally to Muse Spark Open—just replace BASE_MODEL with an identifier from the supported list.

The key steps are roughly:

1. Install the SDK: uv add baseten-loops
2. Create a training client, specifying the base model and rank
3. Submit data, execute forward_backward, optim_step
4. save_weights_for_sampler to save a checkpoint
5. Create a sampling client with the checkpoint, then sample directly
6. When training is done, stop the session with the CLI or REST API

Here are a few details worth noting:

• The tweet specifically mentions "with your reward functions." The Loops docs have a dedicated RL page, indicating it is not limited to SFT and can also run reinforcement-learning-style post-training.
• The saved checkpoint is LoRA weights, and the files are small; for example, in the docs example a checkpoint with rank=16 is only 32MB. There are corresponding APIs for listing, downloading, and deploying.
• Both training and sampling run on GPUs, and billing continues while the session is open. The docs specifically remind you to deactivate after training; otherwise, the backend will keep burning money.
• If you want to save money, you can use the environment variable LOOPS_REUSE_FROM_RUN_ID or ServiceClient(reuse_from_run_id=...) to reuse the previous trainer and sampler and skip cold starts.

If you have clear training data and reward functions, it is worth trying Loops. At least for developers who want to quickly validate model post-training, this workflow removes a lot of infrastructure hassle. The quickstart documentation is here: Loops quickstart