wllama v4.0 ditches wasm mem64 for mem32 (up to 22% faster), 2.8x speedup on Chrome

ngxson · x · 2026-11-23

wllama v4.0.0 is out, bringing up to 2.8x faster inference on Chrome, video input, full-duplex realtime audio, and hybrid CPU + WebGPU support on Chrome and Firefox for running bigger models.

Author ngxson explains a counterintuitive design choice: v4 uses wasm mem32 instead of mem64 because mem32 is up to 22% faster. To work around mem32's 4GB heap limit, the wasm runtime is split into multiple instances linked together via RPC — building on rgerganov's GGML RPC backend, applied in an unexpected place.

Related event: wllama v4.0.0 Boosts In-Browser LLM Inference Up to 2.8x(4 posts)→

Original post →

More from coding & agent

coding & agent channel →