HyperFlow H3 Flash
A slimmed MiniMax-H3 that makes video with sound on a single 24 GB card
Emerged finds new products worth a look, daily — what’s new today
It brings text-to-video and text-to-audio generation on-device to M-series Macs, so you can experiment with a large multimodal model without any cloud.
h3.c, also called h3-metal, is a native inference engine from Salvatore Sanfilippo (antirez) that runs the MiniMax H3 multimodal model locally on Apple Silicon Macs. It targets M3 and M5 Max chips and generates video and audio from text, with conditioning on a first or last frame and reference image, video or audio inputs. It uses Metal GPU acceleration plus quantization, token reduction and layer thinning to fit the model into a Mac's memory at usable speed. It is for people with recent Apple Silicon who want on-device multimodal generation. MIT licensed, active development.
More like this
Edge0
Run large mixture-of-experts models on a Mac by streaming experts from SSD
TideGlass
Calm, audio-reactive nature simulations for your second monitor on macOS
Tonefold
Open-source music sketchpad: AI writes parts into a piano roll you can edit
try-omarchy
Run the Omarchy Linux desktop as a native app on Apple Silicon
Horizon
A Rust client for Minecraft 26.3, audited against the real game