> ## Content Index
> Fetch the complete content index at: https://www.testingcatalog.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Leaks: Google testing math-focused DeepThink V3 model
- URL: https://www.testingcatalog.com/google-deepthink-v3-mathematica-leak/
- Published: 2026-09-16T10:03:23.000Z
- Updated: 2026-09-16T10:03:23.000Z
- Description: A fresh leak indicates Google is developing a dedicated math-reasoning variant of DeepThink V3, featuring unfiltered chain-of-thought traces and high-temperature generation.
- Author: Dwaybe Cubbins
- Tags: Google, AI Rumours, Latest AI News

It looks like Google might be working on a new math-specialized DeepThink V3 model. The details come from a couple of screenshots shared by popular leaker [*Lyra*](https://x.com/lyraxana?ref=testingcatalog.com) on X.

The screenshots show the experimental build named “Mathematica” built on top of DeepThink V3 and also the model's backend configuration which shows it being tuned for heavy calculations and complex symbolic problem-solving.

API data shows the model carries the internal identifier *"models/deepthink-mathematica-tf-raw-thoughts"* and supports a massive one-million-token context window alongside a 65,536-token output limit. *"UNSTABLE\_EXPERIMENTAL"* and includes the label *"Teamfood."* In case you're wondering, that label is Google's jargon for internal employee testing.

> HOLY MOTHER OF MATHEMATICS!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!  
>  
> Google is working on a math-focused variant of its DeepThink model, and its raw thoughts are pretty funny. [pic.twitter.com/0mkukSLlQw](https://t.co/0mkukSLlQw?ref=testingcatalog.com)
> 
> — lyra (@lyraxana) [September 15, 2026](https://x.com/lyraxana/status/2099916458549506315?ref%5Fsrc=twsrc%5Etfw&ref=testingcatalog.com)

The screenshots suggest Google is actively working on the project, and we might see it soon.

The interesting part is the model's internal monologue. As the leaker noted, the raw thoughts are quite funny and unfiltered. While working through a tough polynomial Diophantine equation, the AI started cheering. *"HOLY MOTHER OF MATHEMATICS!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!"* and then added, *"Let's admire the absolute beauty of this equation."* Upon finding a shortcut later, it blurted out, *"OH MY GOD!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!"* in pure excitement.

It's not surprising to see Google working on a math-specialized DeepThink model. The company has already been pushing into hardcore reasoning like we previously saw with the [Olympiad-winning Gemini DeepThink IMO](https://www.testingcatalog.com/early-preview-of-olympiad-winning-gemini-deep-think-imo/) mode which was specifically designed for elite math competitions. 

Google also just launched [Gemini 3.8 Live and Live Extended Thinking](https://www.testingcatalog.com/google-rolled-out-gemini-3-8-live-and-extended-thinking/) for multi-step reasoning.

That said, we'd still take this with a grain of salt. Google has not yet publicly confirmed the new model. It may or may not reach the public, depending on how it performs in internal tests.