Thinking Machines' Inkling scores 36.5% on ARC-AGI-2

Thinking Machines' Inkling model posted 79.5% on ARC-AGI-1 and 36.5% on ARC-AGI-2, per Latent Space's July 19 coverage.

ARC-AGI-2 is the harder of the two benchmarks, the one where every lab's score has stayed low. Latent Space calls 36.5% meaningful movement in a spot where movement has been rare. That's the number enterprise buyers evaluating reasoning models should watch, not the ARC-AGI-1 score.

Red Hat is deploying Inkling on DGX B200 hardware, served through vLLM, the open-source inference server. That puts a reasoning-focused model into a standard enterprise inference stack. It's the same kind of setup enterprise teams already run production inference on, not a research demo.

Yes, but 36.5% still means Inkling misses roughly two of every three ARC-AGI-2 problems.

The coverage doesn't say how this score compares to Thinking Machines' earlier releases. It also doesn't say whether Red Hat's DGX B200 deployment is a pilot or a standing commitment.