Qwen 3.5: Alibaba Challenges ChatGPT with a Powerful Open Source AI

Qwen 3.5: A Significant Advancement in Open Source AI

Alibaba, the Chinese tech giant, has recently introduced Qwen 3.5, an artificial intelligence model that presents a serious alternative to ChatGPT. This fully open-source model is designed to run locally on a desktop PC, even outperforming GPT-5 Nano on many benchmarks. This initiative is part of a broader strategy by Chinese labs, such as Moonshot, MiniMax, and DeepSeek, aiming to compete with American proprietary models by developing high-performance open-source solutions.

An Ambitious Open Source Strategy

The 3.5 version of Qwen, announced in February, offers a range of models from 397 billion parameters in its largest version to a lighter version with 27 billion. Between these extremes, there are intermediate models with 122 and 35 billion parameters. The 35 billion version, built on a MoE architecture, requires over 22 GB of VRAM, making it already resource-intensive. In March, Alibaba expanded this range with four new models: 9B, 4B, 2B, and 0.8B, designed to operate on more modest hardware configurations while competing with recent proprietary models.

A Design for Local Execution

On March 2, Qwen 3.5 was enhanced with these four new models. These versions incorporate a hybrid attention system, combining linear and classical attention to optimize resources without sacrificing response quality. Specifically, in four consecutive processing steps, three use linear attention, which is less computationally demanding, while only one employs classical attention, which is more precise but resource-intensive. This approach significantly reduces the resources required to run the model without compromising the quality of the responses.

All released versions are also natively multimodal. Unlike other models that add a vision encoder afterward, Qwen 3.5 integrates visual understanding from the outset. This means that text, images, and videos are processed within the same neural network, without distinction. However, the model only produces text as output. In terms of context capacity, Qwen 3.5 claims a window of 262,000 tokens natively, equivalent to a 500-page novel processed in one go. It is even possible to push this limit up to one million tokens (about 2 hours of video) with a slight loss of accuracy, using YaRN, a mathematical adjustment technique for context size.

Impressive Performance

Qwen 3.5-9B stands out particularly in multimodal vision and reasoning benchmarks, outperforming OpenAI's GPT-5 Nano and Google's Gemini 2.5 Flash-Lite. In document comprehension, it scores 87.7 compared to 55.9 for OpenAI. It also surpasses OpenAI's GPT-OSS-120B in several text tasks despite its smaller size. For example, in scientific reasoning, Qwen 3.5-9B scores 81.7 against 80.1 for GPT-OSS-120B. In general knowledge, it reaches 82.5 compared to 80.8, and in long context comprehension, it shows 55.2 against 48.2.

However, the model shows limitations in programming, with a score of 65.6 on LiveCodeBench, lower than that of GPT-OSS-120B. Advanced math tasks also reveal a gap favoring larger models, with a score of 83.2 compared to 90.0 for OpenAI's model. In clear terms, for classic uses, document analysis, visual reasoning, multilingual understanding, and agents, Qwen 3.5-9B competes at the highest level. But for high-level coding and competitive mathematics, larger models maintain an edge.

Increased Accessibility

Distributed under the Apache 2.0 license, Qwen 3.5 is completely free and usable for commercial purposes. This accessibility could disrupt the AI market, offering a high-performing, cost-free alternative to dominant American solutions. By providing such a capable model at no cost, Alibaba could redefine the standards of open-source AI.