v1.1.0 - Next-Gen Moat: Double-Buffered Prefetching, INT4 Quantization & DeepSeek Sparse MoE

StreamSpec v1.1.0 introduces 3 major systems engineering breakthroughs: 1. DoubleBufferedPrefetcher: Overlaps Layer L+1 memory bus transfer with Layer L tensor core execution, cutting bus stall time to near-zero. 2. BlockQuantizedLinear: Sub-byte 4-bit block-scaled weight streaming (0.5 bytes/param), reducing SSD/PCIe payload by 73.4% with >0.999 cosine similarity fidelity. 3. DeepSeekSparseMoE: Active Expert Unioning for Mixture-of-Experts, skipping 57.8%-80% of unread routed experts and proving a >1.2x Expert Reuse Factor.

v1.0.0 - Asynchronous Speculative Layer-Streaming

Official Flagship Release of StreamSpec v1.0.0. Enables 70B and DeepSeek-R1 inference on consumer silicon with zero GPU clusters using speculative chunk verification, DeepSeek MLA memory compression, and mathematical rejection sampling.