An 80B MoE (Qwen3-Next, 48 GB GGUF) answering in a Chrome tab: llama.cpp's WebGPU backend, experts paged in from disk
Follow-up to last week's post about streaming MoE experts from disk in a browser tab.
Proof1 community thread
Key points
- This week the same disk tier sits under llama.cpp's own WebGPU backend, so it works for the MoE models llama.cpp already runs, with no per-model engine.
- Model (GGUF) File On the GPU Load from disk Decode
- Qwen3-Next 80B-A3B Q4KM 48.4 GB 4.89 GB 48 s 2.9 tok/s (proof of concept)
- The paging is a small change to ggml-webgpu: a new buffer type, slot-mapped MULMATID shaders, and the residency logic.
Sources (1)
- [1]An 80B MoE (Qwen3-Next, 48 GB GGUF) answering in a Chrome tab: llama.cpp's WebGPU backend, experts paged in from diskr/LocalLLaMA (top, daily) · Oct 13, 02:04 PM
Follow-up to last week's post about streaming MoE experts from disk in a browser tab.
This week the same disk tier sits under llama.cpp's own WebGPU backend, so it works for the MoE models llama.cpp already runs, with no per-model engine.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 13, 2026Horsey challenge | Strata Fork | Qwen 3.8 FN NVFP4
- Oct 13, 2026Update: Converting dense models into Mixture-of-Experts
- Oct 13, 2026SonicBoom: 52 tok/s decode on a laptop RTX 3050 (llama.cpp: 45) with the same quantized kernels, the difference is a graph compiler. No Python.
- Oct 13, 2026Running decision models: Picking quant, VRAM usage
- Oct 13, 2026Fish Audio Drama 3 Technical Report
- Oct 12, 2026strix-llama new branch: prompt processing +21–54% over Vulkan and up to 2.7× over stock ROCm kernel at 128k, same outputs quality
