Qwen 3.5: Alibaba’s Open-Weight AI Is Quietly Challenging GPT-5.4 and Gemini 3.1 (2026 Guide)

While the AI industry’s attention has been locked on the OpenAI-Anthropic-Google triad, Alibaba’s Qwen team has been shipping models at a pace that is quietly redrawing the competitive map. Qwen 3.5, released in March 2026, is the most capable open-weight multimodal model available today — and for many real-world tasks, it is not just close to GPT-5.4 and Gemini 3.1 Pro, it is ahead of them.

This is not a fringe claim from a company trying to generate hype. Qwen 3.5 matches or exceeds frontier proprietary models on abstract reasoning (ARC-AGI), graduate-level science (GPQA), and agentic task completion benchmarks. It processes video inputs up to two hours long — a capability no current OpenAI or Anthropic model offers natively. The full model weights are available on Hugging Face under a permissive license. And if you would rather not self-host, Alibaba’s Qwen Chat and the Dashscope API offer access at pricing that undercuts the US labs significantly.

The question for developers, researchers, and AI practitioners in 2026 is no longer whether Chinese open-weight models are competitive. Qwen 3.5 closes that debate. The question is how to use them effectively — and when they are the right choice over the proprietary alternatives.

What Qwen 3.5 Actually Is

Qwen 3.5 is Alibaba’s fifth-generation large language model family, released in March 2026 as the flagship of the Qwen series. It is a multimodal model — meaning it accepts text, images, audio, and video inputs simultaneously — trained on a dataset Alibaba describes as more than 20 trillion tokens, with particular depth in scientific literature, code, multilingual corpora, and long-form video content.

The model family spans several size variants:

• Qwen 3.5-7B: The lightweight option, designed for edge deployment and resource-constrained environments. Runs on consumer hardware with 16GB RAM.
• Qwen 3.5-32B: The mid-tier, comparable to mid-size frontier models on most benchmarks. The sweet spot for most self-hosting use cases.
• Qwen 3.5-72B: The flagship open-weight release. This is the model that benchmarks against GPT-5.4 and Gemini 3.1 Pro.
• Qwen 3.5-Max: A proprietary closed-weight variant optimized specifically for commercial API deployment, with higher context limits and additional alignment work.

The headline capability distinguishing Qwen 3.5 from every other model in its category is native long-video understanding. Qwen 3.5 can process video inputs up to two hours in duration — following plot threads, extracting specific moments, answering questions about visual events that occur at arbitrary timestamps, and generating structured analysis of video content. This is not a summarization hack over extracted frames; the model processes the full temporal structure of the video at the model level.

Benchmark Performance: Where Qwen 3.5 Stands

Benchmarks in AI carry caveats — training data contamination, evaluation methodology differences, and the gap between synthetic benchmarks and real-world performance are all legitimate concerns. With that caveat stated, the numbers for Qwen 3.5-72B are striking.

Abstract Reasoning: ARC-AGI-2

On the ARC-AGI-2 benchmark, which tests abstract pattern recognition on genuinely novel problems resistant to memorization, Qwen 3.5-72B scores 71.3% without code execution. For context, Gemini 3.1 Pro leads the open-to-public models at 77.1%, and GPT-5.4 scores in the upper 60s on the same evaluation. Qwen 3.5-72B is within 6 percentage points of the current leader — a narrower gap than most analysts expected from an open-weight model at this scale.

PhD-Level Science: GPQA Diamond

GPQA Diamond consists of 448 graduate-level multiple-choice questions in biology, chemistry, and physics, designed by PhD researchers to be resistant to lookup. Qwen 3.5-72B scores 87.4%, placing it second overall among publicly benchmarked models. Gemini 3.1 Pro leads at 94.3%; GPT-5.4 scores approximately 89%. Qwen 3.5-72B is competitive with GPT-5.4 and significantly above Claude Opus 4.6 on this benchmark.

Coding: SWE-Bench Verified

On SWE-Bench Verified — the benchmark measuring a model’s ability to resolve real GitHub issues on real open-source codebases — Qwen 3.5-72B scores 53.1%. Claude Opus 4.6 leads at 80.8%; GPT-5.4 is around 78%. Coding is where Qwen 3.5-72B falls most visibly short of US frontier models, though a 53.1% score is still exceptional for a self-hostable open-weight model.

Multilingual: Where Qwen Leads

On multilingual benchmarks — particularly for Chinese, Arabic, Hindi, Indonesian, and other non-English languages — Qwen 3.5-72B outperforms every major US model. This is not unexpected: Alibaba’s training data has far deeper coverage of these languages than Anthropic or OpenAI, and the performance gap on Chinese-language reasoning tasks specifically is substantial. For applications targeting non-English markets, Qwen 3.5 is frequently the right choice regardless of other benchmark comparisons.