Pure-Rust inference engine doubles llama.cpp speed on aging GPUs

A developer released RustingStrata, a pure-Rust single-binary inference engine with no Python dependency, achieving 160 tok/s decoding Qwen3.6-35B-A3B (Q2KXL, some experts on CPU) on an RTX 3060 12GB — about twice as fast as llama.cpp.

2026-11-08 ~ 2026-11-08 · 2 related posts

• RustingStrata hits 160 tok/s on RTX 3060 running 35B MoE, ~2x faster than llama.cpp — Serious_Scarcity7818 · 2026-11-08

1 near-duplicate retellings: Serious_Scarcity7818