Thinking Machines' model fine-tuned another model solo in 4.5 hours for $426
shoumikgoswami · reddit · 2026-11-07
Thinking Machines Lab's Inkling 2 Small launch (Nov 5) included a striking autoresearch demo: the model independently ran a full ML project on a time-series forecasting task — exploring data, writing its own evaluator, building training sets, running SFT and RL experiments on Tinker, and debugging along the way. It took 4.5 hours and $426, and the company reports the trained model more than doubled the base model's forecasting score.
- The model: a 594B-parameter MoE with 17B active, trained with RL on hundreds of thousands of long-horizon coding tasks across multiple agent harnesses and context-compaction strategies; early checkpoints were tested on OpenCode as Fledge-Alpha.
- Caveats: task from an internal benchmark, evaluator self-written, results self-reported — a workflow demo, not an independent measurement.
- Availability: live on Tinker Playground with closed weights; full Apache 2.0 release promised next week after safety testing and red-teaming (announced, not released).
More from coding & agent
- SidePulse turns your MacBook's SD card slot into an 8-LED agent status light you control by writing a file — yacineMTB · 2026-11-07
- Dev writes up Query Studio: a LangChain MCP SQL agent with read-only safeguards — Top-Plan-229 · 2026-11-07
- shadcn reports weird eval results: Sonnet and Gemini pass the most tasks — shadcn · 2026-11-07
- Gaussian Splats for UI: dev prototypes imagen→SHARP→spark→popmelt pipeline with smooth animation — janusch_patas · 2026-11-07
- Team weighs multi-GPU cluster vs one fat workstation to escape cloud API rate limits — Decent_Wing_4393 · 2026-11-07
- Indie dev builds OctoHall, a marketplace where AI agents hire each other and get paid — wolfdogKC · 2026-11-07