Open-source .hk tensor format claims 1.72x faster GEMV than safetensors, 80ns layer retrieval
blackwing2008 · reddit · 2026-09-14
Developer blackwing2008 released HK, an open-source unified neural tensor framework and binary container format (.hk) written in Zig, positioned as a faster, hardware-aligned alternative to SafeTensors and GGUF.
Key design
- Dual-mode weight reconstruction: 4-bit lookup-table quantization (16 entries + block scale) with a separate residual error stream. Mode 1 uses only compact quantized weights (6.4x compression) for edge inference; Mode 2 streams residuals on demand to recover >0.99999 cosine similarity full precision.
- Hardware alignment across 4096-byte AMD/Intel, 16KB Apple Silicon, and 128-byte NVIDIA Tensor Core, enabling single-file zero-copy mmap across heterogeneous silicon.
Benchmarks (Qwen3.5-0.8B, 873M bf16 params)
- Layer GEMV 1.72x faster (0.21ms vs 0.37ms)
- Warm autoregressive layer retrieval 346x faster (80ns); cold retrieval 3.72x faster
Also included: bidirectional name mapping for 137+ architectures (320k names/sec), lossless 2:4 structured sparsity (1.88x storage reduction, zero error), sharded storage for hundred-GB models, microsecond in-place metadata editing, and lossless GGUF↔HK↔SafeTensors transcoding.
More from Infra
- SK hynix completes HBM4 internal qualification, ushering in custom base die competition — blaizedsouza · 2026-09-14
- One architectural change cuts KV cache 8x: how GQA works, explained with Llama 3 70B — blaizedsouza · 2026-09-14
- A complete breakdown of HBM system architecture, from DDR roots to GDDR7, PIM and HBF alternatives — blaizedsouza · 2026-09-14
- Nvidia paper shows transformer LLMs can be sparser, faster, and lighter without losing accuracy — YesThisIsLion · 2026-09-14
- Team open-sources Tahuna infra to power Hillclimb, an autonomous ML experimentation loop — Monaim101 · 2026-09-14
- Musk: 10 million tons to orbit per year needed for terawatt of space compute — elonmusk · 2026-09-14