Key takeaways

• Qwen3.5 27B is a multimodal vision-language Transformer model designed for fast response times, balancing inference speed and performance, with overall capabilities comparable to Qwen3.5-122B-A10B.
• The model excels in direct factual recall, semantic, and linguistic understanding across diverse domains, demonstrating strong capabilities in mathematical problem-solving, including symbolic algebra, systems of equations, number theory, and geometry.
• It achieved 97.6% accuracy on the MATH-500 dataset, showcasing excellent ability in complex mathematical reasoning.
• The model struggles with complex reasoning, multi-step problem-solving, nuanced contextual interpretation (e.g., legal, historical, biological), and integrating visual interpretation with text, often misinterpreting subtle contextual cues or providing factual inaccuracies in niche knowledge domains.
• It scored only 21.44% accuracy on 'Humanity's Last Exam' (Medical LLM Benchmark), indicating significant weaknesses in clinically relevant contexts, complex reasoning, and niche factual recall. Common failure modes include misinterpreting contextual cues, factual inaccuracies in niche domains, and errors in multi-step reasoning.

Technical specifications

Context length

262k tokens

Max output

—

Architecture

Transformer

Modalities

text+image->text

Parameters

~27B

Open weights

Yes

Licence

Apache 2.0

Regions

China

Vendor

Qwen

Provider model key

qwen/qwen3.5-27b