33B MiniMax-H3-Image image model runs fully on a Snapdragon phone NPU, no cloud
shirosoro · reddit · 2026-11-15
HexAnima 1.6.0 adds MiniMax-H3-Image—a 33B-parameter DiT image model with a Qwen3-VL text encoder—to its catalog, running the full pipeline (text encoder, 50 DiT blocks, VAE) INT8-ConvRot quantized on the Hexagon NPU of a Snapdragon 8 Gen 1 phone. No cloud, no PC, no GPU offload.
- Why it fits: the 48 GB of weights are streamed block-by-block from UFS flash instead of residing in RAM; small images are storage-read bound, large ones NPU-compute bound.
- Benchmarks: 512×512 in 83s, 1024×1024 in 124s, 2048×2048 in 406s, all at 6 steps.
- Storage: 50 GB, downloaded from the public HF repo ayanoby/hexanima-h3 with per-file sha256 manifests and resumable downloads.
- License: MiniMax H3 Community License excludes US/EU/UK/South Korea and bars using outputs to train other models; the app requires consent before downloading.
More from Infra
- srush's blog verifies floating-point certificates for 152 ML operators, finds unreported bugs — ChrSzegedy · 2026-11-15
- AGI would still face a world running on decades-old infrastructure no one dares to touch — SuB8u · 2026-11-15
- SpaceX cuts AI datacenter build time to ~4 months, MiniHard done in 5 — Dr_Singularity · 2026-11-15
- LMDeck: a macOS menu-bar app to unload local LLMs across backends in one click — Fickle_Tradition4491 · 2026-11-15
- Kosmo and Ropedia release Physical AI dataset: 1,589 3D-scanned environments, 131 hours of egocentric video — liuziwei7 · 2026-11-15
- Retired Microsoft engineer explains what broke in the Outlook outage — nicolascraske · 2026-11-15