RiverRider/blackwindow-mlc-libs

Black Window phone model libraries

WebGPU model libraries for WebLLM, compiled for phones. They use the standard mlc-ai/*-q4f16_1-MLC weights unchanged; only the library differs.

WebLLM's prebuilt libraries are compiled with a 1,024-token prefill chunk, and MLC plans the working buffers from that compiled figure, so on Qwen3 about 0.6 GB sits beside the weights whatever the page asks for at runtime. An iPhone tab is killed somewhere between 1.1 and 1.6 GB. These libraries are compiled with prefill_chunk_size=128 and context_window_size=1024:

library	parameters	temporary buffer	at a 1K window
Qwen3-0.6B-q4f16_1_ctx1024_cs128-webgpu.wasm	320 MB	79 MB	~0.5 GB
Qwen3-1.7B-q4f16_1_ctx1024_cs128-webgpu.wasm	923 MB	83 MB	~1.1 GB

Figures are mlc_llm compile's own accounting. Built 2026-09-10 with apache/tvm e85fbb1 (2026-08-18), mlc-llm 9fa644f5 (2026-08-17), emsdk 4.0.23, for WebLLM 0.2.85 (@mlc-ai/web-runtime 0.27). Used by https://blackwindow.xyz on phones.

Use in a WebLLM ModelRecord:

{ model: "https://huggingface.co/mlc-ai/Qwen3-1.7B-q4f16_1-MLC", model_id: "Qwen3-1.7B-q4f16_1-MLC",
  model_lib: "https://huggingface.co/RiverRider/blackwindow-mlc-libs/resolve/main/Qwen3-1.7B-q4f16_1_ctx1024_cs128-webgpu.wasm",
  overrides: { context_window_size: 1024, prefill_chunk_size: 128 } }

Sunstone North Labs LLC · burton@sunstonenorth.com