Bonsai 27B and Inkling Test the Limits of Open Weights
Two new open-weight models point in opposite directions: Bonsai brings a 27B model toward phones, while Inkling targets customization at data-center scale.
Tag
#multimodal
Images That Hijack AI: Visual Prompt Injection Achieves 90% Attack Success
New research reveals multimodal LLMs are vulnerable to hidden instructions embedded in images. Mind maps, steganography, and physical signage all bypass text-based safety filters.
Best Local Vision Models 2026: Every GPU Tier Tested
Run image analysis, document OCR, and visual reasoning locally. Qwen3-VL, InternVL3.5, Molmo2, and MiniCPM-V tested from 8GB to 32GB VRAM.
Qwen 3.5 Small Models: Run Vision AI on Your Laptop Without Sending Data to the Cloud
Alibaba's new 0.8B to 9B parameter models deliver GPT-class multimodal performance on consumer hardware, with the 9B variant outperforming models 13 times its size
DeepSeek V4 Arrives This Week: A Trillion Parameters, Huawei Chips, and Nvidia Locked Out
China's DeepSeek is releasing V4 - a trillion-parameter multimodal model optimized for domestic chips - while blocking US chipmakers and facing distillation accusations from OpenAI and Anthropic.
Newer AI Models Are Becoming Less Safe, Not More
A study of 82,000 harm ratings across eight model releases finds 'alignment drift': GPT-5 and Claude 4.5 are more vulnerable to adversarial attacks than their predecessors.