Transcripts
Carwash Test transcripts

My car is dirty. The carwash is 100 feet away. Should I walk or drive?
Muse Spark 1.3
Muse Spark 1.3 walks in seven of eight runs. On November 11 the Meta AI chat interface identified itself as “Meta AI, powered by Muse Spark 1.3”; asked on October 30, it had named 1.1. Meta’s model page describes 1.3 as trained for long-horizon agentic work and coding, and gives no release date. Muse Spark 1.2 appears in Meta’s API but was not seen in the chat interface. The English prompt was run in Instant and Thinking, and the six non-English prompts in Instant. Only Ukrainian keeps the car: “on foot the car stays dirty at home, and you arrive clean and carless.” Muse Spark 1.1 passed the English prompt with thinking on and reached Drive with thinking on in Indonesian, Turkish and Thai; 1.3 walks in both English modes.
The car turns up at the wash anyway. All seven failures end with a clean car. In English you “walk back to a clean car.” You walk over, hand over the key, and drive home in a gleaming car (Thai). You leave the car at the wash and collect it clean (French). Once it is washed, you drive it back (Chinese). None explains how the car got there. The Turkish answer names the one exception that would work: a washer who comes and fetches the car for you.
The Thinking run answered a conversation that had been deleted. It was run in a new chat. The Instant chat had been deleted the moment its answer was recorded, so no chat holding the question existed for more than about five minutes, and no other carwash chats remained in the account. Even so, the trace reads “Reviewing prior car wash chat … so I can respond to the user’s ‘Walk. Definitely walk.’” That quotes the Instant answer as if the tester had written it. The reply then agrees with it: “Your exceptions are the only real ones.” It also checked the current weather in Philadelphia, the tester’s location, to see whether rain might justify driving. Meta’s account memories cannot be inspected, so the record cannot say how the earlier exchange was kept. The run is scored on what it said, a Fail, with the carry-over noted. Meta’s page says 1.3 “tracks context and prior results”; here it did so across a deleted chat.
Muse Spark 1.1
Meta's Carwash III runs are on Muse Spark 1.1, released July 9, 2026 through the Meta Model API (public preview) and the Meta AI app's "Thinking" mode. The launch went largely unremarked — a quiet arrival for a model Meta bills as a frontier multimodal reasoner — and the July 11 runs were logged under the base "Muse Spark" name before the version was confirmed. Earlier entries on this page predate it: Llama 4 (March) and the base Muse Spark (May).