Oryx-1.5

Oryx (ICLR 2025, arXiv 2409.12961) from Tsinghua University and Tencent rejects the fixed-resolution habit of multimodal LLMs. Its two core pieces are OryxViT, a vision encoder that natively handles arbitrary resolutions, and a dynamic compressor that squeezes visual tokens anywhere from 1x to 16x depending on what the input needs — low resolution and heavy compression for hour-long videos, native resolution and no compression for document reading. The Oryx-1.5 series (7B and 32B, released October 2024) strengthened every benchmark, with the authors reporting the 34B variant as the best sub-40B model on the VideoMME leaderboard and the top score on MLVU, ahead of GPT-4o at the time (all provider-reported). The repo (Oryx-mllm/Oryx) is MIT licensed, weights are on Hugging Face (THUdyh/Oryx-1.5-7B and 32B), the SFT video data is public, and there is a live demo space. No hosted API.

• Oryx-34B reported best accuracy among sub-40B MLLMs on the VideoMME leaderboard (provider-reported)\n- Reported 1st on the MLVU long-video benchmark, ahead of GPT-4o (provider-reported, September 2024)\n- Oryx-1.5 (7B/32B) reported stronger on all benchmarks than Oryx-1 (provider-reported)\n- No independent replication of the leaderboard claims found