Tag: #sota9 results
A mysterious model named '大象' (Daxiang/Elephant) with only 100 billion parameters has achieved state-of-the-art (SOTA) performance. It excels in token efficiency, making it highly promising for efficient AI inference. The news comes from Qubit.
Seedance 2.5 has been officially released, claiming to set a new benchmark for domestic video generation models. The update emphasizes both state-of-the-art performance and cost-effectiveness.
Papers With Code has updated its platform to include performance evaluations for closed-source models like GPT-5.5 and Mythos 5. Users can now toggle between open-source and closed-source leaderboards to better compare state-of-the-art AI performance.
HiDream.ai has launched HiDream-O1-Image-Pro, a 200B+ parameter image model built on the Unified Transformer (UiT) architecture. The model achieves SOTA performance in text rendering, image generation, and multi-task editing by integrating images, text, and conditions into a shared latent space.
Gaode has released ABot, the world's first full-stack embodied technology system designed for AGI. It achieves 15 state-of-the-art (SOTA) results across key benchmarks. The system builds a continuously evolving closed loop for embodied intelligence.
Alibaba's Qwen team launched Qwen3.5-Omni, a full-modal AI model. It supports speech recognition in 113 languages and speech generation in 36 languages. The model sets 215 state-of-the-art results across multimodal benchmarks.
New 4.5B param model ColQwen3.5-v1, based on Qwen3.5-4B with ColPali late-interaction, tops ViDoRe V1 at nDCG@5 0.917. Apache 2.0 on Hugging Face; v2 targets ViDoRe V3 SOTA. Trained via 4 phases with hard negatives and finance doc specialization.
Hugging Face's PapersWithCode has added conference-specific browsing, starting with CVPR 2026. Users can now filter papers by task, GitHub links, Hugging Face artifacts, and presentation type.
Post critiques modern ML PhDs as often incremental: combining ideas, tweaking for benchmarks, claiming SOTA without deeper insights. Questions if top papers yield lasting knowledge and ponders field incentives. Asks what distinguishes strong incremental work.