DiffusionBear v0.5.0 ditches PyTorch for pure-MLX runtime, .app shrinks 52%
Puzzleheaded_Note739 · reddit · 2026-11-16
DiffusionBear v0.5.0 removes PyTorch entirely from its packaged runtime, rebuilding macOS image generation on pure Apple MLX engines: the .app shrank from 1.3GB to 615MB (-52%) and the zip from 396.7MB to 217.1MB. New in this release: a 3/4-stage memory lifecycle, Virtual Boxes, Kromoss 5B support, RGBA transparency for Qwen-Image 2.1, SHA-256 model integrity checks, and an open Agentic REST API. Apple Silicon only (16GB recommended), macOS 26.2+; weights download on first use from HF/Civitai, then generation is fully offline. MIT-licensed, with a 220-run benchmark arena and video tutorial.
More from Infra
- WSJ: Firm tied to Jensen Huang's ex-mistress arranged $12bn in Nvidia chips sent to China — mjdramstead · 2026-11-16
- llama.cpp already supports /v1/realtime for local speech-to-speech chat with LFM3-Omni — ngxson · 2026-11-16
- Lium vs CoreWeave: H100 from $1.85/hr, claims significantly cheaper GPU rental — markjeffrey · 2026-11-16
- Nokia Bell Labs demos real-time trans-oceanic fiber sensing over 4400 km subsea cable via RDMA — jwt0625 · 2026-11-16
- Berth v0.6.0 leases one RTX 4090 across 23 engines, 143 lease-hours a week — Distinct-Pie2389 · 2026-11-16
- RL on heterogeneous hardware: one B300 trainer, four chips, plain HTTP, zero restarts — DavidBennett__ · 2026-11-16