Google's leaked Mathematica shows the raw chain of thought

Two model stories where the interesting part is what the lab is willing to show: Google's internal math model runs on traces nobody has cleaned up, and StepFun finally put a date on open weights.

An experimental Google model codenamed Mathematica has surfaced in leaked evaluation screenshots, and its most revealing detail is the unfiltered chain of thought. The build carries the internal identifier models/deepthink-mathematica-tf-raw-thoughts — the "raw thoughts" suffix is the point — with a one-million-token context window and a 65,536-token output ceiling, roughly eight times what a standard model is allowed to emit in one response. It is tagged UNSTABLE_EXPERIMENTAL and Teamfood, Google's term for a build circulating among employees rather than customers, and the backend configuration shows it tuned for heavy computation and complex symbolic problem-solving on top of DeepThink V3. The screenshots were posted to X by @lyraxana on September 16 and first written up by TestingCatalog; Google has not confirmed the model exists, and it may never ship.

Why it matters is not the arithmetic. Commercial models have their thinking scrubbed before you see it — RLHF alignment, formatting passes, tone polish — so what reaches the screen is a cleaned-up narrative of a search that was never that tidy. The leaked trace reads the other way: on a Diophantine equation it shouts in capitals when it spots a clean algebraic simplification, then returns to the algebra and verifies the identity. The screenshots also reportedly show the model generating at high temperature, the setting labs normally pin near zero for math and code, which means Google is letting it wander the solution space instead of reciting memorized steps.

That is the real signal in a leak full of noise: an unfiltered trace is a supervision tool before it is a product. It is what you need if you want to reward good intermediate steps rather than just the final answer, and it is the same material that makes a model harder to monitor once its reasoning stops looking like English. Treat the rest cautiously — there is no announcement, no model card, and no benchmark, only a screenshot of a configuration screen.

StepFun announced Step 5 Preview on September 20 — and this time the weights have a date: October 15. The Shanghai lab's new flagship is a sparse mixture-of-experts model with 600 billion total parameters and 27 billion active per token, a one-million-token context window, native text and image input, and a price of $1.00 per million input tokens and $2.70 per million output, with a 95% cache discount. The API and AI Studio opened the same day, and a bfloat16 repository has already been reserved under the lab's Hugging Face account for the October release. We covered the leak two days before the announcement — StepFun's Step 5 Preview leaks at 600B — and the weights stay closed — and the open-weights commitment is the part that resolves it.

StepFun is leading with efficiency: at 27 billion active parameters the model bills like something far smaller while carrying a frontier-sized footprint. Two numbers deserve to sit next to that claim. On Artificial Analysis's index the model scored 44, level with Kimi K3, but it generated 160 million output tokens to finish the run against a median near 92 million — verbose, and at $2.70 per million output tokens that verbosity is most of the bill. And a third-party configuration circulating in developer tooling lists the same model at 350,000 tokens of context with a 64,000-token output cap, which contradicts the one-million-token window StepFun advertises; if your pipeline depends on the long context, check the documentation rather than the leaderboard.

What to watch: whether the October 15 checkpoint actually lands, and whether Mathematica shows up publicly with its traces still unfiltered.

Would you rather see a model's raw, messy reasoning or a cleaned-up summary of it? Tell us in the comments.

Sources: TestingCatalog · lyra on X · IT之家 via Sina Finance · 36Kr · StepFun · Artificial Analysis · OrcaRouter · IT之家 via Baidu