HyperFlow: 14B distilled MiniMax H3 runs video gen on 6GB VRAM at 2x speed
cgpixel23 · reddit · 2026-11-19
A hands-on test of HyperFlow, a 14B open-source model created by self-distillation acceleration of the 20B MiniMax H3, packaged with a full ComfyUI workflow:
- Supports NVIDIA Sol-Attn sparse attention, combined with the SelfLift sampler for extra speed, plus RelayVSR upscaling
- No LoRA needed, works at only 6 steps; character and motion handling beat other MiniMax H3 workflows the author tried
- 0.7 megapixels in 20 minutes on 6GB VRAM, roughly 2x faster than the author's previous setup
- Main drawback: distorted faces in some long-distance prompts, still being tuned
- Weights on Hugging Face, nodes and workflow on GitHub/Civitai, video tutorial on YouTube; full workflow on Patreon
More from Infra
- Ex-DeepSeek core member open-sources Linger, a personal AI agent that runs local models on your PC — FellMentKE · 2026-11-19
- German bank's LLM compliance dilemma: DIY Langfuse vs Orq.ai for audit trails — CommonMinimum587 · 2026-11-19
- US blocks Chinese firms from renting Nvidia chips abroad while Microsoft nears $500M OpenAI deal for ByteDance — kevinnbass · 2026-11-19
- vikhyatk: we are overinvesting in GPU compilers, underinvesting in CPU compilers — mohbibi_ · 2026-11-19
- KVARK Flash: a proxy that caps DeepSeek's thinking per request, 30-50% fewer tokens — Ill_Contribution_717 · 2026-11-19
- Polymarket launches Anthropic IPO perps, ANT trades near $104 ahead of Nasdaq debut — Polymarket · 2026-11-19