5 Trillion Context. Zero Tokens.

One 4D field, one shot, about 22 terabytes out. Why that number is a resolution, not a window.

Accelerated Understanding, the company Anima Anandkumar and Benedikt Jenik took out of stealth on August 25, says its physics model has run inference at more than 5 trillion context. Frontier language models handle about a million, sometimes 10 million in experimental settings. That gap is not a bigger window on the same axis.

The model has no tokens. Its context is one physical field, sampled in three dimensions of space plus one of time, and the 5 trillion is how finely that field was sampled. This sheet reads the company's own published page, marks what is drawn to scale, and marks what is still blank.

Two founders, one website. Anandkumar is the Bren Professor of Computing and Mathematical Sciences at Caltech and was previously Senior Director of AI Research at NVIDIA. Jenik built and ran large-scale machine learning systems at online marketplaces, with a background in self-driving cars and early large models at MIT. Their page credits them with FourCastNet (2022), which it calls the first large-scale high-resolution AI weather model.

What they published is a longform page, a scroll explainer, a comparison table, and one video from a sub-100B training run. No paper, no weights, no API, no benchmark table. The news page reads "Coming soon."

The numbers on the page: hundreds of pre-training runs at sizes up to 1 trillion parameters; one scaling test at 35 trillion parameters, which they describe as limited by the cluster available at the time; 2 to 6 petabytes of data in an average training run; up to 1 trillion context in training and beyond 5 trillion at inference, with no patching or sub-sampling; and one inference output at 5 trillion context weighing about 22 terabytes.

Reuters carried the launch, as reported by OfficeChai and Startup Fortune: the model uses neural operators rather than Transformers, the target markets are chip design, robotics, extreme weather prediction and geological data for energy companies, and the founders declined an earlier invitation from the Jeff Bezos-backed Project Prometheus. None of that architecture language appears on the company's own site.

For a language model, context is one-dimensional: a sequence of tokens. For a physics model in their setup, context covers three spatial axes plus time. Their own worked example: 1,000 points along each spatial axis over 1,000 time steps multiplies out to 1 trillion. The 5 trillion figure is the same object at a finer grid, roughly 1,500 per axis and 1,500 steps if you keep it cubic (that last line is arithmetic, not a company figure).

The cost of that choice is stated plainly on the page. A single training sample can be too large for one accelerator or one node, and the same holds for individual model layers, so they left FSDP behind and wrote their own sharding. That is a more concrete engineering claim than most launch pages make, and it is the reason the 5 trillion is not a context-window trick: there is no way to sub-sample your way to 22 terabytes of output.

Direct 4D. The model takes the state of a system and predicts the full time rollout in one pass. They reject two shortcuts: flattening space to 2D frames (loses observation detail) and stepping forward one frame at a time (errors compound). One-shot rollout is why a single sample is so large.

Resolution invariance. The same model runs at any level of detail, in training and at inference, so a design loop can start crude and get finer without switching models. The line about having tested up to 5 trillion context sits under this heading on their page, which is the second clue that the figure is a resolution.

Many kinds of physics in one model. They report that a model trained across physics domains outperforms models trained on each domain alone at equal size, and call it cross-physics uplift. The data argument follows: industrial datasets are sparse where they matter, barely enough to fine-tune a broad model, nowhere near enough to train a narrow surrogate from scratch.

Directional feedback. An experiment returns an outcome. Their model returns an outcome plus a direction of improvement, which can run in a loop against a design goal while staying physically valid.

The one artifact backing the table is the video: a sub-100B run producing full rollouts in one shot across several areas of physics from a single model, played back frame by frame for humans. They note the model had not seen heat sink shapes at training time. It is a demo, and it is the closest thing on the page to evidence.

This is the section worth reading twice. Their argument for why cyber capabilities improved so fast: tight feedback loops (a result in seconds, not a lab cycle) and clean signals (pass or fail). Most fields lack both. Physics, they argue, has something better than both.

The governing equations for most areas of interest are known. Checking whether a candidate field satisfies its equation costs little. Producing a field that satisfies it is the hard part. So the equation becomes a verifier that sits outside the model, and unlike a cyber capture-the-flag, it does not return a bit. In their words, you get "continuous scores and even a gradient that you can optimize for." That signal trains the model past what the data alone would teach, and the same signal steers a design loop: define the task and a baseline, simulate, take a small step in the direction of improvement, repeat.

The idea is not new to this team. Training a neural operator against the residual of a partial differential equation is the mechanism in Physics-Informed Neural Operator (2021), with Anandkumar among the authors. What the launch page adds is the claim that this signal holds up at trillion-parameter scale, across many kinds of physics in one model, and doubles as the optimizer for the thing being designed.

A drawing sheet marks unknown dimensions as blank rather than guessing them. These are blank.

In fairness, the page does not overclaim on the point that matters. It says the models do not have to be flawless from the start, that they need to be good enough to give useful feedback, and that you can still check the result in the lab at the end. That is a smaller promise than replacing simulation, and a more believable one.

Context is a per-modality quantity. Comparing 5 trillion with 1 million compares grid points to tokens. The comparison that would matter is cost per useful answer and whether the answer is right, and neither is on the page.

A verifier that returns a gradient beats one that returns a bit. If your domain has a checkable law, the residual is both a training signal and an inexpensive evaluator, even when generation is expensive. Unit tests give a coding agent a bit. This gives a physics model a slope.

One-shot rollout is a bet against compounding error, paid for in sample size. It is the same trade you see between parallel and autoregressive generation elsewhere. Here the price is a sample that does not fit on one node.

A page with numbers and no baselines is a specification, not a result. Read the 1T, 5T and 22 TB as what the system is built to hold, and wait for the error table before reading them as what it gets right.

The strongest sentence on the page is not the 5 trillion. It is that a lab experiment tells you what happened, and their model claims to tell you which direction to move. Until an error table sits next to that claim, the 22 terabytes is the most concrete thing they have shipped.