On September 18, the source @lyraxana posted tweets on platform X on September 16thGoogleBased on the DeepThink V3 modelBuild an internal experiment codenamed Mathematica AI Models, WHICH IS ESPECIALLY OPTIMIZED FOR SOLVING COMPLEX SYMBOL PROBLEMS, MAINLY IN HEAVY COMPUTING, AND IS EXPECTED TO BE THE BEST MATHEMATICAL REASONING AI MODEL OF GOOGLE。
API data show that the model's internal identifier is “models / deepthink-mathematica-tf-raw-thoughts” and supports context windows up to 1 million characters (token), with an output limit of 65,536。
Note: The context window refers to the total number of text units that can be read while the model is being processed; the output ceiling refers to the number of text units that can be generated by the model's single response。
The leak configuration also labels "UNSTABLE_EXPERIMETAL". This status name indicates that the version is an unstable experimental version. The model is accompanied by the “Teamfood” label, which is the term Google uses for internal employee testing, meaning that the service is within the scope of internal testing。
TestingCatalog points out that Google has previously launched the Gemini DeepThink IMO model for high-level mathematics competitions and the Gemini 3.8 Live and Live Extened Thinking on September 15。