Thinking Machines released Inkling 2 Small. It is a Mixture-of-Experts model with 594B total and 17B active parameters. It supports image inputs and variable effort. It scores 31 on the Artificial Analysis Intelligence Index. Improvements span agentic coding, vision, knowledge work, and reasoning. It is available now on Tinker with closed weights. Full weights will be open-sourced under Apache 2.0 next week after safety testing. It can build websites, games, and 3D designs from scratch. It can also emulate a Linux PC and run a kernel in the browser. It was trained on long-horizon coding and autoresearch tasks. It outperforms many peers on coding benchmarks like SWEBench Pro and Vibe Code Bench. It autonomously trained a time-series model in under five hours. It uses a 292M-parameter vision encoder for image processing.
What changed
Inkling 1 models had lower scores on agentic coding, vision, knowledge work, and reasoning.
What it unlocks
Running long-horizon autonomous coding, vision, and knowledge work tasks in a single model.
- 594B total parameters
- 17B active parameters
- 31 on Artificial Analysis Intelligence Index
- $426 Tinker spend for time-series task
What you need to act on it
- access to Tinker
- closed weights until next week
Sources