StrataFlux preprint reports improved low-bit inference using streamed precision

Researchers presenting StrataFlux report that selectively streaming additional numerical precision from host RAM and solid-state drives can improve compressed language models without surrendering their memory advantages. The preprint describes overlapping transfers with computation and reusing fetched precision across generation steps. For Qwen3-8B with 2-bit weights and key-value cache, the framework raised AIME 2024/2025 accuracy from 11.0% to 47.1% at unchanged GPU memory use. For 2-bit Llama-3.1-8B-Instruct weights on RTX 5090, it reduced WikiText-2 perplexity from 9.46 to 8.40 while retaining 95.4% of the compressed baseline's throughput. The findings come from evaluations of selected models on dedicated GPUs and unified-memory platforms.