FastH3-Omni lands Oct 20 with open weights, local RTX/Spark/MLX support
Vandy_simp · reddit · 2026-10-14
FastH3-Omni, a reference-conditioned video model generating synchronized audio, opens weights on October 20 with local support for RTX GPUs, DGX Spark and Apple Silicon via MLX. It supports character/setting reference images, first-and-last-frame conditioning, up to 9 image references (3 in real-time mode), audio references for voices and soundtracks, dialogue generation, clip continuation, and 5–15s outputs at 24 FPS. API preview is live.
Related event: FastH3-Omni Set for October 20 Open-Source Release(2 posts)→
More from Multimodal
- Gemini 4 Argon turns your webcam feed into a playable synthesizer app — GeminiApp · 2026-10-14
- Strikingly lifelike AI short film imagines Hannibal before he was famous — aeeiee · 2026-10-14
- Prime Intellect acquires Moondream to build efficient multimodal models — Sentdex · 2026-10-14
- Testing Seedance 2.5 on AI video's hardest challenge: keeping a knot consistent — Salmaaboukarr · 2026-10-14
- PS2-graphics LoRA for MiniMax H3 turns AI video into retro cutscenes — aziib · 2026-10-14
- Learning MiniMax H3: forcing fully_copy ref_audio_0 in ref-to-video workflows — Ill-Purchase-3312 · 2026-10-14