Story

On air. The free wire runs 15 minutes behind live. See pricing

World, crypto, finance and tech news, scored for impact.

CryptoTokenPostKorean, translatedFiled 33 h ago

512×512 77.8ms… 122.2 times faster than Jeburation, Janus Pro-7B

Original headline512×512 77.8ms…제버레이션, 야누스 프로-7B보다 122.2배

Open the sourceReported by TokenPost, tokenpost.kr, 19 November 2026, by 김민준

Not investment advice. Muhabbit scores how much a story may matter, not which way to trade.

Researchers at the National University of Singapore have released Jeveration, a unified vision model that generates 512×512 images in 77.8ms, which the paper reports as 122.2× faster than Janus-Pro-7B. The system pairs a frozen Qwen3.5-4B vision-language model with shared LoRA adapters and task-specific output heads, handling text-to-image, image-to-image, and segmentation, depth, and surface-normal prediction from a single checkpoint. Jeveration treats each image as a joint decision over 128 discrete codes, conditioning every position on the same random state so parallel prediction preserves whole-image coherence instead of the sequential token generation used by autoregressive models or the iterative denoising used by diffusion models. The result matters because it points to a single compact backbone replacing separate per-task pipelines while cutting inference latency sharply. What to watch next is whether the reported 77.8ms and 122.2× figures hold up under independent benchmarking and how the shared-adapter approach scales beyond the 4B-parameter base.

Muhabbit's AI wrote this summary in its own words from TokenPost's report and translated it from Korean. It can get details wrong, so check the source.

Named in this story

Jeveration, National University of Singapore, Qwen3.5-4B, Janus Pro-7B