# Modal Releases Open-Source Spindle Training Server

> Modal’s Spindle lets teams run Tinker-compatible model training in their own Modal workspace and pay for GPU time rather than tokens. It is still in beta.

Oossa · 2026-10-22 · https://oossa.com/en/modal-releases-open-source-spindle-training-server

Modal has released Spindle, an open-source, Tinker-compatible training server that runs in a team’s Modal workspace. Users write their own training loop, while Spindle handles distributed training; Modal says teams can customize the server and pay for GPU time instead of per-token API pricing.

Modal reports lower costs than Tinker in several workloads, including a 63% reduction for one DAPO-math setup and 25% for an eight-client search-agent setup. These are Modal’s own comparisons, and the company says Spindle is still in beta.

## The facts

- Spindle supports full-parameter training and multi-LoRA across a range of open models, according to Modal.
- In Modal’s DAPO-math comparison, Spindle cost $3.57 per million output tokens versus Tinker’s $9.77 in the reported setup.

## Why it matters

Teams running repeated model-training experiments can use Spindle to run custom training loops on Modal and may lower costs by sharing GPU capacity; actual savings depend on workload, and Modal’s comparisons are vendor-reported.

## Sources & references

1. [Introducing our Training API](https://modal.com/blog/introducing-training-api) – Modal, 2026-10-22

Last updated: 2026-10-22