Modal releases Spindle, an open-source Tinker-compatible training server

Why Modal built an alternative to Tinker

Modal says Tinker changed how people do post-training: users write the training loop in plain Python on a laptop, and Tinker runs the forward and backward passes on its GPUs. Modal calls itself a big fan, as are many of its customers. But as those teams scale up, Modal says they keep raising two problems.

• Per-token pricing becomes expensive once teams run dozens of experiments.
• Users can only use the models and features the provider chooses to support.

Modal's answer is that a good API deserves more than one implementation. Spindle can be deployed with a single command, spindle deploy, after which users run their training script against the server. Because the customer owns the server, Modal says, anything in it can be customized. The default configs are tuned for GPU utilization, and the company plans to keep shipping optimizations.

Where Spindle fits in Modal's stack

Modal says it now covers the whole post-training stack. Teams that do not want to write a training loop can use Modal Dojo, a training library the company released the day before. Teams that want to own everything down to the cluster can use on-demand multi-node clusters and sandboxes for RL rollouts.

Spindle sits between the two. Users write the loop with their own losses, rewards and environments, and Spindle handles the distributed training. It supports full-parameter training and multi-LoRA on a wide range of open models, and, according to Modal, hundreds of concurrent training runs at similar performance and significantly reduced cost compared with closed-source APIs.

The Tinker-compatible API is backed by a distributed control plane for request and session management, plus container autoscaling built on Modal Functions. The backend uses disaggregated trainer engines and inference pools, which do not have to run on the same accelerators or precisions. Modal Clusters schedule requests for large models on dynamically available RDMA-connected nodes without a cluster reservation, Modal Volumes handle checkpointing, and Modal Servers provide inference. Modal also says it upstreamed improvements to Miles and SGLang while building multi-LoRA training.

Parity tests and customization

Modal reproduced the project Training Search Agents with GRPO by Jasper Lu, which trains gpt-oss-20b to run multi-turn search over SEC filings. The team changed only the BASE_URL and API_KEY to point at Spindle. Modal reports that Spindle tracks both its own Tinker rerun and the author's published results to within 0.01 reward across all updates, and that checkpoints at updates 16, 20 and 24 scored similarly on 32 held-out queries.

As an example of customization, Modal wanted to train Qwen3.5-9B on long-horizon software tasks but only had a default 16k-context config. Two config changes raised it to 256k: setting max_context_length to "256k" and context_parallel_size to 4. The run lasted about 3 days on preemptible GPUs, resuming automatically from checkpoints when machines were reclaimed, with 8 clients sharing the same GPUs through multi-LoRA.

How multi-LoRA lowers cost

Modal explains that agentic RL spends most of its time generating rollouts, so trainer GPUs sit idle while a client waits. Depending on the workload, a trainer serving one client was busy less than half the time. Spindle's multi-LoRA backend lets several clients share a trainer, so one client's training step runs while another waits on rollouts.

In a GRPO job with a 27B model on DAPO-math, using 4×H200s for training and 12×B200s for inference, four interleaving clients kept the trainer about 88% busy. Adding clients lowers the cost per run but slows each run slightly. Past four clients in that setup, Modal observed trainer thrashing that significantly degraded per-client throughput.

Modal's comparison against Tinker, running the same scripts:

• DAPO-math (Qwen3.8-27B): $9.77 per 1M output tokens on Tinker; $3.57 on Spindle with 2 clients (−63%) and $1.99 with 8 clients (−80%).
• Search agent, up to 40 turns (gpt-oss-20b): $5.06 on Tinker; $4.06 on Spindle with 4 clients (−20%) and $3.80 with 8 clients (−25%).
• SWE-Gym, 256K context (Qwen3.8-27B): $10.79 (est.) on Tinker; $6.17 on Spindle with 4 clients (−43%) and $5.82 with 8 clients (−46%).

What it means for teams

Modal's takeaway is that keeping GPUs busy, for example by running several experiments at once, gives Spindle much better price-performance than closed-source APIs. The figures above come from Modal's own tests, not independent benchmarks.

Spindle is open source and still in beta. Users can deploy it in their Modal workspace, open issues for missing models or features, and submit pull requests for new model configs or performance tuning. For larger contributions, Modal invites users to discuss ideas in its Slack. The company says it is working on performance optimizations, improved autoscaling, new features and training recipes.