The company describes it as compatible with Thinking Machines' Tinker API and says it can be deployed into a Modal workspace with spindle deploy.
Spindle is positioned between Modal Dojo, the company's training library, and Modal's lower-level cluster and sandbox primitives. Modal says it supports full-parameter training and multi-LoRA across a range of open models. Its backend separates trainer engines from inference pools and includes multi-client batch scheduling, asynchronous checkpointing, delta weight synchronization, and routing for multi-LoRA workloads.
Modal says users pay for GPU time rather than per-token API usage and can modify the server because it runs in their workspace. The post reports internal RL runs, including a reproduction using gpt-oss-20b and a Qwen3.5-9B run configured for 256k context. It does not establish independent price-performance results, GPU availability, or a complete list of supported models and pricing.