Qwen3-VL-4B

Headlines · 3

1. 7d
   Open-o3-Video-v2 brings agentic evidence seeking to video reasoning

   Open-o3-Video-v2, an extension of Open-o3-Video, formulates video reasoning as a unified agentic, evidence-grounded process: the model decides what information is missing, actively acquires extra observations, writes the supporting moments and object-level bounding boxes into its trajectory, then continues or answers. To support this behavior the authors built STGR-v2-CoT135K, containing 52K multi-turn agentic grounded reasoning trajectories, plus STGR-v2-RL-13K for reinforcement learning, and propose Clue-Seek GRPO (CS-GRPO), which tracks how tool calls change accumulated evidence, rewards clue coverage and step-wise gains, and penalizes excessive, repetitive or uninformative retrieval. On V-STaR, Open-o3-Video-v2-4B reaches 40.4% mAM and 56.3% mLGM, beating its Qwen3-VL-4B backbone by 8.5% and 12.6% respectively, with consistent gains on general video understanding, long-video reasoning and temporal grounding benchmarks; code and models are released.

   Hugging Face · Papers · 🔥 0
2. Oct 8
   MiniMax H3 live video generation on a single RTX 5090

   A Comfy blog post presents ComfyStreamerH3, an open-source custom ComfyUI node that runs MiniMax H3 on one 32GB RTX 5090, producing 15 seconds of 448×256 video in 15 seconds or less with a four-step setup. The listed optimizations include sparse attention that applies full attention to only about 20% of video blocks, NicoLab28's ClipProj adapting Qwen3-VL-4B as a stand-in for H3's 32B text encoder (about 4.5GB vs 15.7GB), FastVideo's pruned INT8 checkpoint, a fused FP4 MLP, and Kijai's INT8 tiled decoder overlapped with NVENC encoding, cutting VRAM needs from 80GB to under 30GB. The author says running such models continuously can cost upwards of $6,000 per day while RTX 5090 rentals start near 70 cents an hour, and notes FastVideo's Preview V1 took 47.2 seconds for a 15-second 1344×768 clip on one B200. The clips still show visible artifacts, and the author suggests passing them through a cheap upscaler such as SolRefinery.

   Comfy Blog · 🔥 0
3. Oct 7
   SpaceCast-Bench: Evaluating predictive spatial reasoning in VLMs

   The paper introduces SpaceCast-Bench, which the authors describe as the first benchmark to directly and diagnostically evaluate predictive spatial reasoning, where a model must construct a scene from observations, anticipate how an intervention changes it, and reason about the unseen outcome. Built around an observe-transform-infer framework, it contains 3,862 questions from 182 real-world scenes across 16 task types at three levels: static perception, local prediction, and global prediction. Evaluating 21 models shows the strongest reaching only 58.0% versus 87.2% human performance, while spatially specialized models remain near random chance; fine-tuning on programmatically generated data lifts Qwen3-VL-4B from 34.0% to 65.7% with macro-average gains on six out-of-domain benchmarks.

   Hugging Face · Papers · 🔥 0

Experience and discussion from the community

Share my Qwen3-VL-4B experienceAsk about Qwen3-VL-4B