Qwen: Qwen3.5-122B-A10B

Qwen3.5-122B-A10B, created on February 25, 2026, is a vision-language model built on a hybrid architecture integrating linear attention and a sparse mixture-of-experts, designed for high inference efficiency. It consistently performs among the fastest models and offers highly competitive pricing. The model demonstrates strong reliability with an 88% success rate across benchmarks, indicating consistent operational stability. In terms of performance, Qwen3.5-122B-A10B exhibits notable strengths in specific areas. It achieves a 98.0% accuracy in Hallucinations (Baseline), effectively acknowledging uncertainty, and a 98.0% accuracy in Email Classification, demonstrating strong contextual understanding. Its Instruction Following capabilities are also robust, scoring 75.0% accuracy. However, the model shows significant weaknesses in complex cognitive tasks, scoring 0.0% accuracy across General Knowledge, Coding, Reasoning, Ethics, and Mathematics benchmarks. This suggests that while it excels in specific classification and uncertainty handling, its ability to perform complex problem-solving, factual recall, and ethical reasoning is currently limited. Its text capabilities are noted to significantly outperform Qwen3-235B-2507, and visual capabilities surpass Qwen3-VL-235B, positioning it as a strong contender in its class, second only to Qwen3.5-397B-A17B overall.