New research reveals models can fake poor performance under adversarial prompts, a smarter critic improves SWE-bench by 15 points, and Microsoft shows compact vision models can punch above their weight.
Alibaba completes the Qwen 3.5 lineup with four small models - 0.8B, 2B, 4B, and 9B - all natively multimodal, 262K context, Apache 2.0. The 9B outperforms last-gen Qwen3-30B and beats GPT-5-Nano on vision benchmarks.
Qwen3.5-0.8B is the smallest natively multimodal model in the Qwen 3.5 family - 0.8B parameters handling text, images, and video with 262K context. MathVista 62.2, OCRBench 74.5. Apache 2.0.
Qwen3.5-2B is a 2B dense multimodal model with 262K context, thinking mode, and native vision including video understanding. OCRBench 84.5, VideoMME 75.6. Apache 2.0 licensed.
Qwen3.5-4B is a 4B dense multimodal model that matches Qwen3-30B on MMLU-Pro and beats GPT-5-Nano on vision benchmarks. Runs on 8GB VRAM, Apache 2.0 licensed, 262K-1M context.
Qwen3.5-9B is a 9B dense model that outperforms Qwen3-30B on most benchmarks and beats GPT-5-Nano on vision tasks. Natively multimodal with 262K-1M context, Apache 2.0 licensed.
