Thinking Machines Lab is sharing how it trains a large model with lower-precision arithmetic—an engineering advance that helps get more Super Intelligence (SI) research out of the hardware already in use.
American Super Intelligence IndexPublished · News & analysis
Making the numbers work harder.
One route to better SI tools is bigger computers. Another is making each computer do more useful work. A new engineering report from Thinking Machines Lab takes the second route: carefully reducing the precision of the numbers used during model training.
In End‑to‑End 4‑Bit Training, published November 13, the lab describes the recipe behind Inkling 2 Small. It reports an 8% wall-clock speedup in pretraining and a 20% speedup in reinforcement learning in its largest jobs, compared with its full MXFP8 training setup.
These are the lab’s reported training results on NVIDIA GB300 hardware, not ASIndex model scores or a promise that every model will see the same gain.
A carefully mixed recipe.
Training depends on enormous numbers of calculations. Representing values with fewer bits can reduce memory use and accelerate arithmetic, but it also introduces rounding error. The trick is deciding where that trade is worthwhile.
The team uses NVIDIA’s NVFP4, a four-bit floating-point format with additional scaling information, in the expert layers’ forward passes—the calculations that produce a model’s outputs. For the backward passes that help update what the model has learned, it uses MXFP8, an eight-bit microscaling format.
A technique called stochastic rounding sometimes rounds a value up and sometimes down, rather than always choosing the nearest number. The team uses it while converting between representations to reduce systematic bias. Some layers remain at higher precision.
There is a practical lesson beyond arithmetic, too: once calculations became faster, moving data between graphics processing units became a bottleneck. The researchers also adjusted how that data is represented during communication, helping the machines spend less time waiting.
What the team reports.
The speed gain is not presented as free. In pretraining, the team reports about a 2% cost in compute efficiency: compared with MXFP8, its recipe needed somewhat more arithmetic work to reach the same training-loss level, while still delivering the reported wall-clock speedup.
For reinforcement learning, which improves a model through feedback on generated work, the team reports faster end-to-end training with no discernible difference in reward. These findings describe its own setup and measurements.
The report shares the engineering choices behind those results. It is a technical write-up, not an announcement that the model’s weights have been made available to everyone.
Why efficiency matters.
Our reading: this is a useful example of progress below the chatbot interface. Better numerical methods can help researchers run experiments sooner and make more of existing hardware, without every advance depending on another data center.
The wider benefit is a set of concrete ideas other engineers can investigate. Lower precision is not a magic shortcut; it is a tool that becomes valuable when the arithmetic, software and hardware are designed to work together.