Modal open-sources Spindle, a Tinker-compatible training server you pay for in GPU time
andersonbcdefg · x · 2026-10-23
Thinking Machines' Tinker changed post-training: write the training loop in plain Python locally, Tinker runs forward/backward passes on its GPUs. But teams scaling up kept hitting two problems — per-token pricing gets expensive across dozens of experiments, and you're limited to the models and features the provider supports.
Modal's answer: Spindle, an open-source, drop-in Tinker-compatible training server that runs in your own Modal workspace. You own the server, can customize anything, and pay for GPU time instead of tokens. Default configs are tuned to maximize GPU utilization out of the box, with ongoing price-performance optimizations planned. Deployment takes a single spindle deploy command, then point your training script at the server.
Related event: Modal Open-Sources Spindle, a Tinker-Compatible Self-Hosted Training Server(2 posts)→
More from coding & agent
- Testing Hermes Agent with n8n: Letting Agents Execute Real Automation Workflows — techlatest_net · 2026-10-23
- OpenAI: Keep MCP Server Instructions Under ~1,000 Characters or ChatGPT May Truncate Them — bibryam · 2026-10-23
- Only ~56% of MCP Servers Provide Instructions, OpenAI Data Shows — bibryam · 2026-10-23
- OpenAI: Don't Hide MCP Tools Behind listCommand/executeCommand Dispatchers — bibryam · 2026-10-23
- Resend logs 1M monthly MCP calls as OpenAI frames MCP as a distribution channel — bibryam · 2026-10-23
- MCP Elicitation lands in Claude: servers can ask users questions mid-action — bibryam · 2026-10-23