Applying NSN-Based 3-Bit KV Compression to Looped Transformers — Evaluating Actual Storage, Prediction Quality, and Fused Readout Costs

Can a looped Transformer use a smaller KV cache without losing the time it saves to reconstructing the stored values? We integrated NSN-based 3-bit compression into Ouro-1.4B and Nanbeige4.2-3B-Base, keeping model weights fixed. The implementation tracks keys and values by recurrence and layer, stores actual packed 3-bit codes together with reconstruction metadata, and uses the cache for subsequent predictions. We evaluated four texts per model using teacher forcing: 255 predictions per text for Ouro, and 8 then 256 predictions on the same four texts for Nanbeige. The fused path met the prespecified 5% diagnostic threshold for relative perplexity degradation in every designated window. Final KV storage, including metadata and the uncompressed tail, fell by 67.68% for Ouro and by 64.39% in the extended Nanbeige evaluation. To isolate the cost of materializing a full-length V array, we compared two paths that used the same packed values and the same reconstruction arithmetic. Reconstructing each block inside attention and using it directly in the weighted sum reduced component time by 28.28%. In Nanbeige, initial compression and four subsequent compression updates offset most of the forward-pass savings; total time over 256 steps was approximately equal to the checked uncompressed route. This companion paper provides the implementation and measurement details for the looped-model experiments in the integrated NSN study (DOI: 10.5281/zenodo.23938878). Readable code, experiment records and reproduction instructions: https://github.com/metaSATOKEN/nsn.