inclusionAI releases ZwZ2 models (27B/9B/4B) based on Qwen3.6-27B

·2026.10.30 20:17

Key point

The new models use full-image inference without zooming tools, improving overall accuracy from 55.52 to 60.51 across six domains.

inclusionAI released ZwZ2-27B, 9B, and 4B models, with the 27B variant based on Qwen3.6-27B.
Overall accuracy increased from 55.52 to 60.51 (+4.99 points) across 25 benchmarks in six domains, outperforming the base model on every benchmark.
The method uses EvidenceConnect to generate verified multi-region questions from SA-1B images and trains with GRPO using approximately 20K examples.
Inference requires no zooming tools, relying instead on full images with colored region outlines to connect visual evidence across regions.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.