Quantizing every part of a large language model to 4-bit still matches full-precision performance.

Big language models use lots of memory. Researchers often compress them using quantization, which shrinks the numbers that represent model weights down to fewer bits. A popular model combines regular attention with recurrent layers called Gated DeltaNet. People feared that shrinking these recurrent layers to 4-bit precision would make errors snowball over long texts, so they left them larger. Researchers tested this idea by building Minima. Minima compresses all 496 linear layers down to NVFP4 4-bit precision, including every recurrent layer. It works because tiny sixteen-element blocks handle extreme outlier values, and the math functions that protect the memory gates squeeze down calculation errors. Meanwhile, the recurrent rules actively overwrite old noise over time instead of letting it build up. Minima runs faster, uses only 17.5 GiB of memory, and matches the original full-precision model's performance on tests almost completely.

Why it matters

This makes large AI models smaller and faster to run, cutting hardware costs and enabling use on more devices. It shows aggressive quantization doesn't harm accuracy, so AI can be deployed more widely.

Questions answered