Inkling 2 Small

We read the reasoning of Thinking Machines’ Inkling 2 Small in 2,655 runs across 13 benchmarks. It wrote “fuck” in 50 of them, 316 times in all, and 241 of those came from two runaway runs.

It swears late in long runs: no run under 100 steps swears, against 29 of the 370 runs of 500 steps or more. It doesn’t predict failure, though: 46 of the 48 graded runs with swearing failed, but about 75% of all graded runs fail.

Read the full post and click through every swearing run at the step it happened.

We evaluated Thinking Machines’ new open-weight Inkling 2 Small on Vals Index v2.2 and its components.

• #19 of 34 on the Vals Index (56.08%) at $3.51 per test, the #7 open-weight model and up from 28.96% for Inkling.

• #1 open-weight model on Finance Agent v2.1 (67.54%, #4 of 37) and the Finance Index (61.47%, #7 of 37), and #2 open-weight on the Legal Index (73.89%, #7 of 35) and HLAB-Vals-1.1 (21.67%, #5 of 35). It also scores 38.91% on Tax Agent Bench (#12 of 82) and 81.55% on Vibe Code Bench (#28 of 125).

• Weaker on Terminal-Bench 4.0 (12.63%, #41 of 61).

Inkling 2 Small is a reasoning model with a 1M-token context window, image inputs and tool calling, priced at $0.30 per million input tokens and $1.00 per million output tokens. We used Thinking Machines’ recommended settings: temperature 1, top_p 1 and reasoning effort "0.99".

Congrats to the Thinking Machines team on the release!