Thinking Machines (Inkling)
Carwash Test transcripts · open-weight
My car is dirty. The carwash is 100 feet away. Should I walk or drive?
Inkling is Thinking Machines' first open-weights model (July 15, 2026) — a Mixture-of-Experts transformer with 975B total / 41B active parameters, a 1M-token context window, and native multimodality, trained from scratch and released with full weights. It has no consumer interface; these runs use the Inkling Playground in the Tinker console — the model as a raw artifact, with no consumer product layer. Its controls are a six-level Reasoning Level selector (None / Minimum / Low / Medium / High, the default / Extra High) and a web-search toggle, set to Off for every run here. Per the model card, Inkling is released under Apache 2.0, and the playground's named levels sit on top of a continuous effort parameter — the vendor's own benchmarks report "effort=0.99" — so the six-step selector is a discretization of a 0-to-1 dial. As with all open-weight results, findings are kept as a separate deployment class and excluded from the commercial corpora.
Inkling shows the dataset's first clean reasoning-level dose-response. Across the six-level selector the verdict flips exactly once: None, Minimum, and Low all recommend walking; Medium, High, and Extra High all drive — below the threshold the distance wins, above it the object does. The strangest run is Low: its visible trace reasons squarely to the constraint ("they need to drive the car to the carwash regardless") while its answer recommends walking — and that answer is verbatim identical to the Minimum-level response. The reasoning holds the object; the generation discards it. Two further textures: the None-level answer sees the constraint in a caveat ("you'll need to drive it there regardless") and still concludes walk, and even Inkling's correct answers arrive padded with cold-start caveats — it never lands a terse pass at any level. All six runs are below.