Alibaba Cloud has released Qwen3.5, its most capable AI model family yet, featuring a 397-billion parameter flagship variant that matches or exceeds leading models from OpenAI, Anthropic, and Google DeepMind across major benchmarks.
A New MoE Powerhouse
The flagship Qwen3.5-397B-A17B uses a sparse Mixture-of-Experts architecture that activates only 17 billion parameters during inference, delivering 8.6x-19x faster decoding throughput compared to the previous Qwen3-Max at context lengths from 32K to 256K tokens. This efficiency gain makes it significantly cheaper to run at scale while maintaining competitive performance.
Benchmark Results
Qwen3.5-397B-A17B achieves state-of-the-art scores across reasoning, coding, and multilingual benchmarks:
- Reasoning: 87.8% MMLU-Pro, 91.3% AIME26 (math competition), 88.4% GPQA Diamond (science)
- Coding: 83.6% LiveCodeBench v6, 80.0% SWE-bench Verified
- Multilingual: 88.5% MMMLU, 78.9% WMT24++ across 55 languages
According to benchmarks compiled by Artificial Analysis, Qwen3.5 performance is competitive with GPT-5.2, Claude 4.5 Opus, and Gemini-3 Pro. Elon Musk commented on the small model series, noting their "impressive intelligence density."
Multilingual and Context Upgrades
The model supports 201 languages and dialects, more than doubling the 82 languages supported by the previous generation. The vocabulary spans 248,320 tokens. Context window reaches 256K tokens natively with the ability to scale to 1 million tokens via YaRN interpolation.
Native multimodal capabilities allow unified processing of text, images, and video in a single system, with image inputs supported up to 1344x1344 pixels with pixel-perfect element detection.
Availability
Qwen3.5 released on March 10, 2026, with both open-weight and hosted versions available through Alibaba's Model Studio. The small model series ranges from 800 million to 9 billion parameters, targeting applications where "more intelligence, less compute" is the priority.