ZoomGain-4B: a vision-language model that learns from what zooming reveals
InclusionAI quietly shipped ZoomGain-4B (and 9B), a Qwen3.5-based vision-language model trained with reinforcement learning on questions built from details that only appear when you zoom into an image, plus a 1,000-question ZoomGainBench and the full RL training dataset.
Why it matters
Most VLMs miss the small print: ZoomGain mines the gap between what a model sees at full-image scale and what a zoomed crop reveals, then trains on exactly that gap — and it shipped with weights, benchmark and data and no announcement at all.
What you could build with it
A team building a visual inspection tool could fine-tune ZoomGain-4B on their own defect photos, using its zoom-based questions to catch small flaws a whole-image model misses, then run the 1,000-question ZoomGainBench to measure the gain.
Does it hold up?
Too early to judge: weights, benchmark and training data all landed 11 November with no independent evaluation and no hosted endpoint; the paper's own benchmark is built by the same pipeline as the training data.
Built with ZoomGain-4B
- ZoomGain-4B shipped quietly with no announcement behind itblog · Timestamp-by-timestamp reconstruction of the silent release: weights on 10 Nov, paper and benchmark on 11 Nov, no launch post anywhere.
- ZoomGain: Mining Global-to-Zoom Perception Gaps for Fine-Grained Visual Supervisionarticle · The arXiv paper behind the model: the global-to-zoom perception-gap pipeline, the 1,000-question ZoomGainBench where the best of 28 VLMs scores 61.08, and the 9K-sample RL training that lifts Qwen3.5-4B/9B on held-out perception benchmarks.
- inclusionAI/ZoomGain-4Bhuggingface · The model card for the 4B checkpoint: RL-trained on Qwen3.5-4B, download and usage instructions, and links to the paper, benchmark, RL data and the 9B sibling.
Learn more
First spotted on hf: source.
More AI models
VoiceStudio: open-source, fully-local ElevenLabs alternative for voice cloning, dubbing, and audiobooksdebpalash/VoiceStudio is a desktop app that runs voice cloning, voice design, video dubbing, dictation, transcription…models · JEV 0.75Edge0-145B: open 145B MoE that runs on a single consumer GPUEdge0 open-sourced 145B and 120B latent-MoE models derived from Qwen3.8-Flash-Next, plus an inference framework that…models · JEV 0.74Strata: run a 125B model on a gaming PCFree, MIT-licensed local inference engine that runs Qwen3.8-Flash-Next (125B MoE, about 6B active per token) on…models · JEV 0.73Google AI Edge Gallery: offline speech recognition and synthesis on-deviceGoogle added fully offline speech recognition and text-to-speech to AI Edge Gallery, with ML Kit APIs on Android and…models · JEV 0.73Soup: fine-tune an 8B LLM on a 4 GB laptop GPU from one YAML fileAn open-source CLI that fine-tunes LLMs from a single YAML file, with layer streaming that trains an 8B model on a 4…models · JEV 0.73Tencent Hunyuan Hy3: open 295B MoE with 21B active under Apache 2.0Tencent open-sourced Hy3, a 295B-parameter mixture-of-experts model with only 21B parameters active per token, under…models · JEV 0.73
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.