Wringer compresses a 4B model's transformer weights to 2.655 bits each

Wringer compresses the transformer weights in Agents-A1-4B to 2.655 bits per weight. The September 12 experiment reports 93.7% mean score retention across IFEval, HumanEval and GSM8K in its original one-round evaluation, relative to the BF16 model.

The compressed figure excludes embeddings and other unquantized parts. Including those, the evaluated language model occupies 4.69 bits per weight overall. Wringer adds a temporary adapter to recover performance, then folds its changes into the compressed weights; no adapter remains at inference.

The current implementation expands the stored weights back to BF16 before running them in vLLM. It reduces stored weight size without demonstrating lower inference-memory requirements or faster generation. The experiment used a 96 GB RTX Pro 6000 Blackwell.

The author reports variation across repeated HumanEval runs and some overlap between calibration data and HumanEval tasks. A check excluding the flagged tasks did not reduce the reported retention, but the reused benchmarks and small experiment still limit broader quality claims.