Mage-Flow W4A8 quantized release runs in 8.9 GB VRAM with ARA
ThaJedi · reddit · 2026-10-13
A Reddit user released a W4A8 4-bit build of microsoft/Mage-Flow running on ComfyUI's native quantized kernels, with an accuracy recovery adapter (ARA) restoring quality toward BF16, plus an optional INT8 text encoder that saves another 3.7 GB — total 8.9 GB VRAM.
Related event: Mage-Flow W4A8 Quantized Build Runs Near-BF16 Quality on 8.9GB VRAM(2 posts)→
More from Multimodal
- Fish Audio releases open-weight TTS model Drama 3 alongside $52M seed round — bigaiguy · 2026-10-14
- Gemini 4 turns green-screen footage into a virtual set via Blender + Nano Banana — bilawalsidhu · 2026-10-14
- User's 'golden formula' prompt on updated Muse gets the model to remix sounds — astvatsaturn · 2026-10-14
- Video editor vibe-codes ComfyUI nodes that cut AI video degradation in half — roychodraws · 2026-10-14
- Gemini 4 Argon builds an Elden Ring-style isometric voxel demo in one prompt — kastnerkyle · 2026-10-14
- Hollywood sold the dream of anyone making movies—until AI made it real — VraserX · 2026-10-14