Deep Dive into Phosphene 3.7.0: Faster H3 Decoding & 48GB Mac Quantization
cocktailpeanut · x · 2026-08-12
Phosphene 3.7.0 detailed its update logs for the Hailuo H3 video model, focusing on local workflows and resource optimization:
- Faster draft decoding: Replaces the full 36-layer decoder with a tiny fast decoder, rendering a 5s draft in about 1 minute.
- True 10s single pass: Supports up to 1344x768 native single-pass 10s generation, eliminating seams from stitched clips.
- CivitAI LoRAs support: New LoRA picker allows one-click import and lossless conversion of ComfyUI character adapters.
- 48GB Mac support: Uses a Q8 quantized engine to drop peak RAM from 43 GiB to 27 GiB with only 5% quality loss, enabling deployment on 48GB Macs.
Related event: Phosphene 3.7.0 Slashes Memory for MiniMax H3 on Mac(2 posts)→
More from Infra
- Transformers.js Surpasses 10 Million Monthly Downloads, Rapid Growth Continues — nicodotdev · 2026-08-12
- d-Matrix Chip Claims 20x Speedup for Qwen Inference — TheKanter · 2026-08-12
- Starlink Offers Free Service in Colombia After Earthquake Until September 12 — DimaZeniuk · 2026-08-12
- Popcorn Open-Sources World's Largest Verified ML Kernel Library After 1.3M+ Benchmarks — aryaman2020 · 2026-08-12
- AI Data Center Investments Drive Local Housing and Job Growth — NinaDSchick · 2026-08-12
- AI Data Centers Projected to Consume Over 1 Trillion Liters of Water Annually by 2028 — kpness · 2026-08-12