Leviathan adds .shn5 lossless BF16 weights format

The author introduces .shn5, a new lossless container format for BF16 and F16 model weights that achieves about 1.5x smaller file sizes by entropy-coding exponents using rANS. Built into the local inference engine Leviathan, the format enables running large mixture-of-experts models like Qwen3-30B-A3B on a 16 GB GPU by streaming compressed experts over PCIe and decoding them on the fly. Benchmarks show the decoder delivers high throughput and matches or beats expert-offload modes in llama.cpp while remaining bit-for-bit identical. Open article →