Qianwen 3.5's four consecutive releases of small models have reached a new high in intelligent density, detonating edge AI.

On March 3rd, it was reported that Alibaba open-sourced four small-sized models of Qianwen 3.5 last night, including Qwen3.5-0.8B/2B/4B/9B. Thanks to the innovation and breakthrough in model technology, the performance of these small-sized Qianwen 3.5 models is extremely powerful: the comprehensive performance of Qwen3.5-9B is comparable to that of models with 10 times more parameters; the Agent ability of Qwen3.5-4B outperforms some current international mainstream models, making it suitable as a multi-modal base for lightweight Agents; Qwen3.5-0.8B/2B are small in size and fast in speed, especially suitable for deployment on devices such as mobile phones and smart glasses. After the open-source release, Elon Musk immediately stated on social media that Qianwen 3.5 has an "impressive intelligence density".

Relying on the innovation in model architecture and breakthrough in training, Qianwen 3.5 has for the first time achieved powerful native multi-modal capabilities in small-sized dense models. Both the intelligence level and visual understanding ability of the model have been enhanced. Small-sized models now also possess the performance level of medium and even large models, reaching a new high in intelligence density. In multiple authoritative evaluations including Instruction Following (IFBench), Doctor-level Reasoning (GPQA), Mathematical Reasoning (HMMT 25), Embodied Reasoning (ERQA), and Complex Document Understanding (OmniDocBench), the performance of Qwen3.5-9B is comparable to that of models like Qwen3-Next-80B-A3B-Thinking, which has 10 times more parameters, and significantly outperforms international mainstream lightweight models. It is a highly cost-effective choice for general-purpose models.

The smaller 4B model perfectly balances performance and resource consumption, with extremely strong Agent capabilities, making it suitable as a multi-modal base for lightweight Agents. In the Visual Agent (ScreenSpot pro) evaluation, the performance of Qwen3.5-4B is on par with that of Qwen3-VL-30B-A3B, which is nearly 8 times larger in size. It can operate mobile phones and computers autonomously like a real person. In the Tool Invocation (TIRE-Bench) evaluation, the performance of the Qianwen 4B model significantly outperforms current international mainstream models. The ultra-small Qwen3.5-0.8B/2B models are small in size and fast in inference speed, and can be directly deployed on terminal hardware such as mobile phones, tablets, smart cockpits, and wearable devices, opening up new possibilities for edge-side AI applications such as offline voice interaction, local document parsing, and real-time perception and decision-making. Some experts said that with the emergence of the small models of Qianwen 3.5, the core AI application scenarios in the future will fully explode on the edge side.

Currently, Alibaba has open-sourced 8 new models of Qianwen 3.5, all of which have achieved "winning with small size", which is also the "intelligence density" improvement that amazed Elon Musk, providing more powerful intelligence with less computing power. The new-generation native multi-modal base model Qwen3.5-397B-A17B released on Chinese New Year's Eve, with less than 400 billion parameters, surpasses the previous generation's flagship model of Qianwen 3 with trillions of parameters. The three medium-sized models Qwen3.5-35B-A3B, Qwen3.5-122B-A10B, and Qwen3.5-27B open-sourced at the end of last month have strong performance and can run on consumer-grade graphics cards. In the first month of its open-source release, Qianwen 3.5 dominated the global open-source model ranking list, occupying four of the top five positions and causing a sensation in the AI community. Some developers found through actual tests that a medium-sized Qianwen 3.5 model can run locally at high speed on an ordinary laptop equipped with an M4 chip, with performance comparable to that of top models. Some developers exclaimed, "Qianwen has single-handedly put a model of Claude Sonnet 4.5 level into the computer for free."

It is understood that Alibaba adheres to full-scale open-source across all sizes and modalities, covering different fields such as large language models, mathematics, programming, voice, and vision. A total of more than 400 Qianwen models have been open-sourced, with a global download volume exceeding 1 billion times and more than 200,000 derivative models. It is a globally influential open-source model system.