Large language models often use low‑bit quantization to save memory and speed. That can hurt accuracy, however. StrataFlux challenges the trade‑off by adding extra precision from the host’s RAM or SSD. The system streams selective refinement bits during inference. It selects which channels or tokens need higher precision and delivers them from external memory while overlapping with GPU computation.

On an RTX 5090, the technique was tested with 2‑bit Llama‑3.1‑8B‑Instruct weights. The original 2‑bit model had a WikiText‑2 perplexity of 9.46. With StrataFlux, perplexity dropped to 8.40. Throughput stayed at 95.4 % of the base quantized model.

Using the same 2‑bit weights together with 2‑bit KV‑cache, the framework raised Qwen3‑8B’s AIME 2024/2025 accuracy from 11.0 % to 47.1 %. GPU memory use did not increase.

StrataFlux relies on usable secondary storage. The refinement bits must be stored in host RAM or an SSD. Without such external memory, the accuracy gains would not be possible.