One year of Tinker: Thinking Machines adds full fine-tuning and inference

Thinking Machines' 15 October look back at Tinker, its fine-tuning API for open-weight models, lists a year of additions. It has gone from 14 to 34+ models (4B to 1T parameters, now with image and audio), added enterprise security controls, full fine-tuning in beta, and OpenAI- and Anthropic-compatible inference for your trained checkpoints, also in beta. It also highlights customers such as Bridgewater, Glean and Sundial training task-specific models that it says beat frontier LLMs at a fraction of the cost. Fact or hype: Partly fact. The features are real; the "beats frontier models" results are customer-reported and narrow.

Why it mattersIt's a strong signal that fine-tuning your own model on your own expertise is becoming practical for non-giant companies, and that you can now serve those models from the same place. But the cost and accuracy wins are self-reported on specific tasks, you still need ML skills to write the training loop, and the high-throughput serving product is waitlist-only. We haven't tested it hands-on.

Curated by Sara Simeone as part of our human editorial coverage. The original sources are linked below; curation is distinct from a Hype or Fact? assessment.

SHARE THISPost on X ↗Share on LinkedIn ↗Copy link

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.