Thinking Machines Lab released Inkling 2 Small, a 594B-parameter Mixture-of-Experts model that activates only 17B parameters per token, routing each token to 8 of 384 experts across 48 layers with a 292M vision encoder for image input. It scores 31 on the Artificial Analysis Intelligence Index and 86.3% on SWE-Bench Pro v2, but is slow per task and weak on Terminal-Bench at 13.1%; weights remain closed with an Apache 2.0 release planned after safety testing.
Visual Encoder
视觉编码器
Core component extracting image features in multimodal models
Curated Intelligence
Researchers introduced RAVEL, a training-free evaluation method that utilizes stabilized PCA whitening and patch-level similarity scoring to address the performance degradation of existing cross-modal metrics in large candidate pools. Achieving a high Spearman correlation of 0.85 without additional training, RAVEL significantly outperforms traditional methods, offering a reliable baseline for efficient vision encoder selection in multimodal LLMs.
Zhipu AI has released and open-sourced GLM-OCR, a lightweight model with only 0.9B parameters that achieves state-of-the-art performance (94.6 score) on OmniDocBench V1.5. Optimized for complex scenarios like handwriting, tables, and code, it supports efficient deployment via vLLM and SGLang, offering a high-performance solution for edge computing and high-concurrency applications.