Physics-IQ Leaderboard: FLUX 3 Leads by 14.7pp, Sora 2 Near Bottom
bfl_ai · x · 2026-10-06
BFL shared the full Physics-IQ Verified leaderboard details: FLUX 3 [large] tops the chart with +14.68pp net improvement, ahead of NVIDIA's open-source Cosmos3 family (Physis-Lang Super/Nano, Cosmos3 Super), Seedance 2.5 and MiniMax H3. Google's Veo 3.1 variants, Alibaba's Wan 2.2 and Tencent's Hunyuan Video 1.5 land mid-pack, while OpenAI's Sora 2 sits near the bottom at -9.93pp. The benchmark covers 31 models from 15 labs, with code open-sourced.
Related event: FLUX 3 Tops DeepMind's Physics-IQ Benchmark, Beating Veo and Sora 2(3 posts)→
More from Multimodal
- Creator shares 120-prompt workflow building an AI video game with Seedance 2.5, Godot and Magnific — techhalla · 2026-10-06
- Paper-cutout comedy short made entirely with Gemini Omni video generation — carinkemo · 2026-10-06
- TikTok 5.6B-Video Dataset Trending on Hugging Face — datasocial · 2026-10-06
- DEPICT: Training-Free Alignment Metric Boosts Negation Accuracy from 19% to 88% — swordhealth · 2026-10-06
- IDU: Unified Unlearning and One-Step Distillation for Flow and Diffusion Models — Aleksei Leonov · 2026-10-06
- Qwen Releases OmniReasoning: Audio-Visual Joint Reasoning Benchmark and 30B Model — Qwen · 2026-10-06