Google reportedly tests Mathematica model for complex math reasoning

Source @lyraxana posted on X that Google is building an experimental AI model internally codenamed Mathematica, based on the DeepThink V3 model. It is mainly optimized for heavy computation and complex symbolic problem solving, and is expected to become Google's strongest math reasoning AI model.

API data shows the model's internal identifier is "models/deepthink-mathematica-tf-raw-thoughts", supporting a context window of up to 1 million tokens with an output limit of 65,536 tokens. The leaked configuration is marked "UNSTABLE_EXPERIMENTAL", indicating it is an unstable experimental version. The model also carries a "Teamfood" tag, Google's term for internal employee testing.

TestingCatalog notes that Google previously launched the Gemini DeepThink IMO mode for advanced math competitions, and recently introduced Gemini 3.8 Live and Live Extended Thinking.