inclusionAI open-sources ZwZ2 vision models: +4.99 points across 25 benchmarks without zoom tools
jacek2023 · reddit · 2026-10-30
inclusionAI released ZwZ2 (27B/9B/4B, Qwen-based finetunes) using Zooming-without-Zooming-v2: models learn to connect visual evidence across regions via comparison, aggregation, matching and multi-step reasoning, with no zooming tools at inference.
Method
- EvidenceConnect builds verified multi-region questions from SA-1B images using privileged zoomed views, filtered via cross-model agreement and drop-one-region tests
- Trained with GRPO directly (no SFT stage), only 20K examples per size
Results: improvements on all 25 benchmarks across six domains, overall accuracy 55.52 → 60.51 (+4.99), spanning fine-grained perception, spatial reasoning, long-context, multimodal agents. GGUF weights and the RL-VQA dataset are public.
More from Multimodal
- Mischief 1: video model born from real brain cells, cheapest image-to-video at 0.5¢/s — _akhaliq · 2026-10-30
- FlexUMM: Unified Multimodal Models Aren't So Unified in Embedding Space — jindong_wang92 · 2026-10-30
- Veo-generated AI short 'Mizuki': from a rooftop to the edge of space — EitherCartographer63 · 2026-10-30
- AI video reimagines Ghostface as a rapper with near-photoreal quality — Easy_Message1869 · 2026-10-30
- Veteran gamedev shows AI-built worm-swarm monster in UE5, says AI raises the ceiling — ID_R_McGregor · 2026-10-30
- AI-generated pulp-noir illustration series nails red-and-black thriller aesthetics — LudovicCreator · 2026-10-30