Deep Think Mathematica is the internal name on a Google model that was never meant to be seen outside the company, and on 15 September 2026 two screenshots put it in front of a million people. The leaker, who posts on X as lyra (@lyraxana), published a raw API model descriptor alongside an unredacted chain-of-thought trace in which the model shouts “HOLY MOTHER OF MATHEMATICS!!!” at a polynomial Diophantine equation. TestingCatalog wrote it up the following morning.
Most of the coverage has focused on the funny bits, which are genuinely funny. The more useful part is the descriptor itself, because it is a machine-readable object with a version string, a model stage, generation methods, sampling defaults and token limits. Read carefully, it tells you what stage Deep Think Mathematica is at, who currently has access, what it is allowed to do, and — just as importantly — which of the claims circulating about it are not actually in the file.
Below we take the leak apart field by field, decode the identifier, explain what “Teamfood” means inside Google, work out what the maths in the second screenshot actually is, set Deep Think Mathematica against the public Gemini Deep Think that anyone can already use, and separate what the screenshots show from what has been read into them.
Table of contents
- What the Deep Think Mathematica Leak Contains
- Reading the Deep Think Mathematica Identifier Piece by Piece
- What UNSTABLE_EXPERIMENTAL and Teamfood Actually Mean
- Why the Deep Think Mathematica Raw Thoughts Look So Strange
- The Problem in the Screenshot Is a Famous Open One
- Deep Think Mathematica Against the Public Deep Think
- What Deep Think Mathematica Would Change in Practice
- Where the Coverage Goes Beyond the Evidence
- How the Deep Think Mathematica Leak Travelled
- What to Watch For Next on Deep Think Mathematica
- References
What the Deep Think Mathematica Leak Contains
The first screenshot is a single JSON object of the kind the Gemini API returns from a model listing endpoint. Every field is worth reading.
The descriptor in full
| Field | Value | What it tells you |
|---|---|---|
| name | models/deepthink-mathematica-tf-raw-thoughts | Internal build, not a product SKU |
| version | gdm-eval-model-u | A DeepMind evaluation slot, letter-indexed |
| displayName / description | Mathematica (DeepThinkV3) Teamfood Raw Thoughts | Both fields carry the same string |
| inputTokenLimit | 1,048,576 | Exactly 2^20, a one-million-token window |
| outputTokenLimit | 65,536 | 2^16, one sixteenth of the input budget |
| supportedGenerationMethods | generateContent, countTokens | No streaming, no live or bidirectional mode |
| temperature / topP / topK | 1 / 0.95 / 64 | Standard Gemini sampling defaults |
| maxTemperature | 2 | The ceiling, not the setting |
| modelStage | UNSTABLE_EXPERIMENTAL | Pre-preview; no stability guarantee |
| thinking | true | Reasoning tokens are produced and exposed |
The two methods are the real signal
Deep Think Mathematica supports generateContent and countTokens, and nothing else. There is no streamGenerateContent and no bidiGenerateContent. That is the profile of an evaluation endpoint someone submits a problem to and waits on, not a conversational product. It is also consistent with a model that thinks for a long time before it says anything. If Deep Think Mathematica routinely spends minutes on a single problem, streaming would be close to useless anyway.
Nothing in the file names a benchmark
There is no accuracy figure, no benchmark tag, no training description and no parameter count. Anything you read about how good Deep Think Mathematica is at mathematics is inference from its name and from one trace, not from the descriptor. That distinction matters more than usual here, because the name is doing an enormous amount of rhetorical work in the coverage.
Reading the Deep Think Mathematica Identifier Piece by Piece
The name field is doing most of the work, and Google’s internal naming is unusually legible here.
deepthink
The family. Deep Think is Google’s high-compute reasoning mode, built on top of the flagship Gemini Pro model and using parallel thinking to explore several solution branches at once. The public version today is Gemini 3.1 Deep Think, and Deep Think Mathematica inherits that lineage rather than starting a new one.
mathematica
The specialisation, and the part that makes this a story. Note that this is an internal codename, not a reference to Wolfram’s Mathematica software, and the descriptor contains nothing that connects the two. Deep Think Mathematica appears to be a mathematics-tuned branch rather than a general model with a maths prompt.
DeepThinkV3
Carried in the display name rather than the identifier. Google has never shipped anything publicly called DeepThink V3, so this is the internal generation counter running ahead of the consumer naming — the same pattern as every other lab’s internal builds.
tf
Teamfood. Google’s internal ladder runs roughly from dogfood, meaning the team that built it, through teamfood, meaning the wider product organisation, to trusted tester and then public preview. The tf in the Deep Think Mathematica identifier places it on the second rung.
raw-thoughts
The most interesting suffix. Public Gemini endpoints return a summarised version of the model’s reasoning. This build returns the unedited stream, which is exactly why the second screenshot reads the way it does. A -raw-thoughts variant is an internal diagnostic tool, and its existence implies a matching non-raw build.
What UNSTABLE_EXPERIMENTAL and Teamfood Actually Mean
Two fields between them set realistic expectations, and both point the same way.
It sits below preview, not above it
UNSTABLE_EXPERIMENTAL is the earliest stage Google exposes in a model descriptor. Models at that stage are rebuilt, renamed and deleted without notice, and the stage field is the single most sobering line in the whole Deep Think Mathematica descriptor. A Deep Think Mathematica build visible today may not exist next week under that name.
Teamfood is a real gate, not a formality
The teamfood rung means Googlers outside the core team can use it and file feedback. Plenty of models die there. Teamfood exposure is why a leak of this kind is possible at all: the population with access to Deep Think Mathematica has grown from a handful of researchers to a large internal audience.
The evaluation slot is telling
gdm-eval-model-u looks like a sequential evaluation-slot identifier rather than a release version. Models sitting in a lettered evaluation slot are usually being scored against a battery, not prepared for a launch date. Deep Think Mathematica is being measured, not packaged.
What none of this rules out
Google shipped an IMO gold-medal Deep Think variant as a limited preview within weeks of announcing the result in 2025. A specialised internal build reaching a narrow external audience quickly is well precedented at this company, even from an unstable stage.
Why the Deep Think Mathematica Raw Thoughts Look So Strange
The second screenshot is where the story went viral, and it deserves a more careful reading than “the AI got excited”.
What the trace actually says
Working through a polynomial Diophantine equation, the model writes “HOLY MOTHER OF MATHEMATICS!!!”, then “LOOK AT THIS POLYNOMIAL DIOPHANTINE EQUATION!!!” and “Let’s admire the absolute beauty of this equation.” Later, on spotting a cancellation, it produces “OH MY GOD!!!” followed by dozens of exclamation marks, then “WAIT!!! LET’S CHECK IF THIS IS JUST THE IDENTITY WE STARTED WITH!!!”
Summarised thoughts hide this by design
Nothing here is unique to Deep Think Mathematica. Every reasoning model produces a scratchpad in a register nobody optimised for readability. What is unusual is seeing it unsummarised, because the public API deliberately post-processes reasoning into something calmer and shorter before returning it.
The affect is a training artefact, not a feeling
Enthusiastic self-talk in a reasoning trace correlates with the human mathematical writing these models learn from, and reinforcement learning on correct solutions can amplify tokens that precede successful steps. Reading emotional states into a Deep Think Mathematica trace is a category error, however entertaining the output is. The exclamation marks are tokens that happened to correlate with progress during training.
It also shows the model self-checking
The most substantive detail is buried in the excitement. The trace repeatedly stops to verify: it checks a simplification, then asks whether the result has collapsed back to the identity it started from. That verify-then-doubt loop is the behaviour the Deep Think architecture is built around, and the Deep Think Mathematica raw trace makes it visible in a way no summarised output ever does.
The Problem in the Screenshot Is a Famous Open One
The maths in the trace is not a toy exercise, and identifying it changes how you read the Deep Think Mathematica leak entirely.
Following the substitutions
The trace sets $S = a+b$, $D = |a-b|$ and $c = 3m – S$, computes $U$ and $V$ in terms of the original variables, and lands on $X + Y + Z = 2(a+b+c) – 18m$. Then it writes the line that identifies the problem: “We know $114 = a^3+b^3+c^3 = (a+b+c)^3 – 3(a+b+c)(ab+bc+ca) + 3abc$.”
That target number is not arbitrary
The equation $x^3 + y^3 + z^3 = n$ over the integers is the sum of three cubes problem. After 33 and 42 were resolved by Booker and Sutherland in 2019, 114 became the smallest value of n for which no solution is known and none has been ruled out. It is one of a short list still open below 1,000.
So the model was attacking an unsolved problem
Whoever was driving Deep Think Mathematica pointed it at the smallest open case of a celebrated number-theory problem. The trace shows it substituting $a+b+c = 3m$ into the symmetric identity to reach $27m^3 – 9m(ab+bc+ca) + 3abc = 114$, which is a legitimate line of attack rather than a hallucinated one.
It does not show a solution
The screenshot ends mid-derivation. Nothing in the leak claims Deep Think Mathematica solved anything, and a genuine solution to the 114 case would be a mathematics news story of its own, not a screenshot on X. What the trace demonstrates is the difficulty class the model is being evaluated at.
Deep Think Mathematica Against the Public Deep Think
The useful comparison for Deep Think Mathematica is not against GPT or Claude. It is against the Deep Think you can already use.
Same limits, different job
| Property | Gemini 3.1 Deep Think (public) | Deep Think Mathematica (leaked) |
|---|---|---|
| Availability | Gemini app, Ultra tier; API early access | Google employees, teamfood stage |
| Scope | General reasoning across science and code | Mathematics-specialised branch |
| Reasoning output | Summarised thought traces | Raw, unsummarised traces |
| Input window | One million tokens | 1,048,576 tokens |
| Tool use | Code execution and Google Search | Not indicated in the descriptor |
| Stability | Generally available | UNSTABLE_EXPERIMENTAL |
The bar the public model already clears
Why a specialised branch makes sense anyway
A general model tuned for chemistry, physics, code and mathematics at once has to compromise. Google already demonstrated what an unconstrained maths variant can do when an advanced Deep Think solved five of six International Mathematical Olympiad problems for 35 points and a gold-medal standard in 2025. Deep Think Mathematica looks like the continuation of that line rather than a new idea.
The precedent for how it would ship
If it ships at all, the likely route is a narrow preview for researchers rather than a button in the Gemini app, which is how the olympiad build was handled. Our coverage of the Gemini 3.8 Live rollout shows the contrasting pattern for consumer-facing models.
What Deep Think Mathematica Would Change in Practice
Assume for a moment that the build ships in some form. It is worth being concrete about who it would help and who it would not.
Research mathematicians
The obvious audience. A model that can hold an entire paper, its references and a formalisation attempt inside a one-million-token window, then reason for as long as it needs, is a literature-search and proof-sketch tool rather than an answer machine. Deep Think Mathematica at that job is a collaborator that never gets bored of checking a substitution.
Formal verification work
The trace’s habit of asking whether a simplification has collapsed back into the original identity is exactly the discipline formal proof assistants enforce. A maths-specialised model paired with Lean or Coq is a more plausible product than a chatbot, and it is a pairing Google has already demonstrated interest in.
Ordinary users, barely at all
Nothing about Deep Think Mathematica suggests a consumer feature. There is no streaming, no tool declaration, no multimodal input and no indication of a latency budget. Homework help is not what an evaluation-slot model with raw reasoning output is for.
Enterprise maths workloads
| Use case | Fit with the leaked profile | Blocker |
|---|---|---|
| Proof search and formalisation | Strong | No stability guarantee at this stage |
| Quantitative modelling review | Plausible | No tool or code-execution field |
| Interactive tutoring | Poor | No streaming or bidirectional method |
| High-volume batch scoring | Unclear | Long thinking time, 65,536-token output cap |
Read that table against the descriptor rather than against the excitement. Two of the four rows are blocked by fields that are actually in the file, which is a more useful way to reason about an unreleased model than guessing at benchmark scores.
Where the Coverage Goes Beyond the Evidence
Three claims have travelled with this leak that the screenshots do not support, and one of them is in the write-up that broke it.
“High-temperature generation”
TestingCatalog’s standfirst describes Deep Think Mathematica as “featuring unfiltered chain-of-thought traces and high-temperature generation”. The descriptor sets temperature to 1, which is the ordinary Gemini default, with a maxTemperature ceiling of 2. A default is not a high setting, and the ceiling is a capability rather than a configuration.
“Tuned for heavy calculations and complex symbolic problem-solving”
The same article says the backend configuration shows this. It does not. The description field is a verbatim copy of displayName, and there is no tuning, objective or training field anywhere in the object. The inference is reasonable from the name; it is not in the configuration.
“A massive one-million-token context window”
Accurate, but not distinguishing. A one-million-token input window is standard across the Gemini Pro line, so the figure says nothing specific about Deep Think Mathematica. The output limit of 65,536 tokens is likewise the family default.
What the leak does support
That an internal maths-specialised DeepThink build exists, that it is at teamfood stage under an evaluation slot, that it emits raw reasoning, and that somebody has pointed it at a genuinely open number-theory problem. That is a solid, narrow set of facts about Deep Think Mathematica, and it is enough to be worth writing about without inflating it.
How the Deep Think Mathematica Leak Travelled
The distribution pattern is worth a paragraph because it explains why a two-image post became a news cycle.
The numbers on the post
A small account, a large reach
The lyra account had 8,405 followers, 141 posts and a May 2026 join date at the time of the leak. A million views on an account that size is platform amplification of the subject, not the author’s audience, and it is the reason a one-line internal descriptor reached a general readership.
The gap before the write-up
The post went up at 17:41 UTC on 15 September. TestingCatalog published at 10:03 UTC on 16 September, a gap of roughly sixteen hours and twenty minutes. Google has made no statement in that window, and none is likely for an unstable internal build. The company communicates on its own schedule, as our write-up of the Gemini Notebook voice mode rollout illustrates.
Treat it as a leak, not an announcement
There is no Google confirmation that Deep Think Mathematica exists, no roadmap entry and no commitment of any kind. Screenshots of API descriptors have been accurate in the past and have also shown models that never shipped. The appropriate confidence is high on existence, low on everything else.
What to Watch For Next on Deep Think Mathematica
If Deep Think Mathematica is heading anywhere beyond an evaluation slot, a few specific things would show it, and none of them is a benchmark rumour.
A non-raw sibling in the same slot
The -raw-thoughts suffix implies a production sibling build with summarised reasoning. A deepthink-mathematica identifier without the suffix appearing in any listing would be a much stronger signal than this leak.
A change of model stage
Movement from UNSTABLE_EXPERIMENTAL to a preview stage, or the appearance of streamGenerateContent in the supported methods, would mark the transition from evaluation to product.
A mathematics result attributed to it
Google has published research collaborations where Deep Think contributed to open problems in algorithms, information theory and combinatorial optimisation. A paper crediting a maths-specialised model would surface the capability before any product page does. Our piece on why mathematicians want proof about AI training data covers the credit questions that follow.
The naming reconciliation
DeepThink V3 does not line up with the public Gemini 3.1 Deep Think branding. Watching how Google reconciles the internal generation counter with the consumer version number will say more about the release plan than the benchmark rumours do.
References
Leaks: Google testing math-focused DeepThink V3 model (TestingCatalog)
Original leak post by lyra on X, 15 September 2026
Gemini 3.1 Deep Think (Google DeepMind)
Gemini with Deep Think achieves gold-medal standard at the IMO (Google DeepMind)
Accelerating mathematical and scientific discovery with Gemini Deep Think
Sums of three cubes (Wikipedia)
Gemini API model reference and token limits (Google AI for Developers)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.