Thinking Machines Lab introduced Inkling 2 Small on November 5, opening its second model generation with text and image input, adjustable reasoning effort and access through Tinker. The mixture-of-experts model has 594 billion total parameters, with 17 billion active per token, according to its official model card.
Visual feedback for longer coding workflows
The focus is longer-running work: coding agents that use visual feedback to build and refine websites, games and interactive designs. Thinking Machines reports improvements over its previous models; we have not independently tested those claims.
Tinker access first, weights planned later
Importantly, available does not mean open-weight yet. The launch provides Tinker Playground, fine-tuning and sampling access. The lab plans an Apache 2.0 weight release the following week, after external safety testing and red-teaming, as described in the launch announcement.
Why it matters
- Developers: another model to evaluate for coding workflows that involve screenshots and visual corrections.
- Small teams like SunMarc: a bounded landing-page task would offer a practical comparison with existing tools.
- Deployment: downloadable weights remain a future milestone, and the lab acknowledges imperfect tool selection—not dependable unattended operation.
For the earlier generation’s focus on customization, see our previous coverage of Thinking Machines’ Inkling models.