Modal launches Spindle: fine-tuning costs fell by up to 80% in its tests

A team is training an agent to write code: it receives a task, edits files, runs tests, and returns for evaluation. Throughout that process, the accelerators allocated to updating the model may be waiting for the next batch of data. The rental bill keeps growing. Modal proposes filling those gaps with other experiments: its new Spindle server lets multiple runs share hardware. According to the company's own tests, costs fell by 20–80% compared with Tinker, depending on the workload and configuration.

This is a useful tool for teams that regularly fine-tune open models and compare many variants. But the central question is how much work they can fit onto the GPUs they are paying for. Switching from per-token pricing to hourly billing does not, by itself, guarantee savings.

How multiple experiments share the same GPUs

Spindle is an open-source server with an API compatible with Thinking Machines' Tinker. Developers define the training loop in Python: how to calculate loss, what to reward the agent for, and which environment it should use to perform tasks. The server, deployed in the team's Modal workspace, handles distributed training. Billing is based on GPU time.

Imagine a team testing four ways to evaluate a search agent's responses. Each variant needs to find documents, read them, and put together an answer. While the first run is busy searching, the server performs a training step for the second. Then the third gets its turn. One experiment's pauses become another's working time.

To enable this, Spindle supports multi-LoRA: multiple runs share a base model while using separate trainable adapters. Full-parameter training is also available. Computation for generating responses is separated from computation for updating weights, so teams can choose different accelerators for each and scale their resources independently.

Compatibility was demonstrated with a search agent built on gpt-oss-20b that works with financial reports and search and reading tools. According to Modal, migrating the existing script from Tinker required only changing the server address and access key. This is a useful migration example, though it does not establish compatibility with every piece of user code.

Where savings emerged and where capacity ran out

In the math test, a single client left long gaps between training steps. When four clients began interleaving their work, training server utilization reached roughly 88%. Adding more clients in this configuration noticeably reduced throughput for each run as they competed for shared resources.

The range of prices illustrates how heavily the results depend on the workload. For math, Modal reduced the estimated cost per million output tokens from $9.77 to $1.99 with eight clients. Savings for the search agent were more modest: from $5.06 to $3.80. For SWE-Gym coding tasks with a 256,000-token context, the company reported a reduction from $15.04 to $5.82.

These are vendor measurements, using different hardware and numbers of concurrent clients. They should not be treated as a promise of a discount for every project. Lower costs also came with longer completion times for individual runs. For a team with a firm deadline for producing one model, that trade-off may not be worthwhile.

What to check before migrating training

Spindle's practical appeal is the ability to adapt the server to a workload. For example, Modal describes extending a model's context from 16,000 to 256,000 tokens through configuration changes and increased context parallelism. However, changing two settings in an example does not mean long context comes free: suitable GPUs and stability testing are still necessary.

A good starting point is to migrate one reproducible experiment. First, compare quality on the same set of tasks, then run several variants concurrently and measure the cost of completed training, waiting time, and accelerator utilization. Cheap tokens are of little help if the result arrives too late.

Spindle is still in beta. Its open source code allows teams to modify the server, but the deployment described uses Modal's infrastructure. For a team with a steady queue of experiments, this is a reason to test the economics of shared GPU utilization. For occasional individual runs, the advantage has yet to be demonstrated. Configurations and results are published in Modal's announcement.