Training-free 20B model Viveki generates real-time video heatmaps under 150ms
not-ekalabya · reddit · 2026-10-22
A Reddit user released the inference method for Viveki, a 20B multimodal decision model (HF: not-ekalabya/viveki-0.1-20B), which generates heatmaps for arbitrary objects or events with <150ms latency on a single A100-80GB, enabling real-time video processing.
The author claims Viveki was built without any training or fine-tuning, based on a mechanistic insight that predicts the model's answers to objective questions in its context before any tokens are generated. Details are intentionally withheld pending a paper. The demo image shows real-time factory CCTV monitoring, answering "is a forklift carrying a tall stack of crates" with 86.9% confidence plus visualized heatmaps.
More from Multimodal
- Kokon 0.10 reuses Qwen-Image 2.1's text encoder to write prompts, runs on 8GB cards — Ok-Director4174 · 2026-10-22
- Speechify launches Simba 3.5: one streaming-native TTS model across 48 languages — baseten · 2026-10-22
- Designer Khyati Trehan uses Gemini 4 Argon 3D for personal illustration work — alexanderchen · 2026-10-22
- Full 3:26 music video generated locally on a single AMD RX 7800 XT with open-source tools — wilppuse · 2026-10-22
- Fictional Lumea Serum ad shows how convincing AI-generated commercials have become — Amaan3024 · 2026-10-22
- Magnific Ships Blender Plugin: Generate 3D Assets and PBR Materials from Your Viewport — aziz4ai · 2026-10-22