Apple releases LensVLM-9B, a vision-language model that selectively expands compressed text images

SEER is a new framework that reduces the computational cost of long-context reasoning in LLMs by selectively compressing text into images, focusing on query-relevant content. It addresses the inefficiency of uniform compression methods by using visual scanning to identify relevant images and retrieve only necessary text, potentially improving both efficiency and accuracy.

Researchers propose TIGER, a text-conditioned visual gated routing framework for multimodal speculative decoding. It dynamically selects relevant visual features to improve token acceptance rates in vision-language models, addressing the limited speedup of existing speculative decoding methods for VLMs.

Apple announced that Mac users in China can now connect to Alibaba's Qwen AI service, expanding AI accessibility in the region. The move integrates Qwen into Apple's ecosystem for Chinese users.

Apple has reportedly trained a China-specific large language model with development support from Alibaba, shifting from its earlier plan to use Alibaba's Qwen for Apple Intelligence in China. Reuters reports Apple is the first foreign company approved by the Chinese government to offer a proprietary AI model in the country, with rollout expected in the coming months. The move aims to navigate China's regulatory environment and strengthen Apple's competitive position.

Two open-source tools, VisionBridge and LocalEyes, enable text-only reasoning models (DeepSeek, Qwen, GLM, CodeLlama) to process images by querying a separate vision model. VisionBridge is an OpenAI-compatible proxy that uses tool calls (look, OCR, scan, crop, compare) and works with LM Studio, Ollama, vLLM, and others. LocalEyes integrates with Claude Code and uses a local Ollama vision model for screen capture, clipboard reading, and image analysis, all without cloud uploads. Both require no training or weight modifications.