Alibaba's New Open-Source Qwen3.5-9B Model Outperforms OpenAI's GPT-OSS-120B and Runs on Standard Laptops
Alibaba's Qwen Team has unveiled its new Qwen3.5 Small Model Series, featuring open-source language and multimodal AI models. Notably, the Qwen3.5-9B, a compact reasoning model, has demonstrated superior performance against OpenAI's 13.5x larger gpt-oss-120B on key third-party benchmarks, including multilingual knowledge and graduate-level reasoning. This series also includes Qwen3.5-0.8B & 2B, optimized for edge devices, and Qwen3.5-4B, a multimodal base for lightweight agents with a 262,144 token context window. These models are significantly smaller than flagship models from OpenAI, Anthropic, and Google, making them comparable to MIT's LiquidAI LFM2 series. The models are available globally under Apache 2.0 licenses on Hugging Face and ModelScope, suitable for enterprise and commercial use. Their technical foundation utilizes an Efficient Hybrid Architecture combining Gated Delta Networks with sparse Mixture-of-Experts (MoE) to overcome memory limitations.
Despite political turmoil in the U.S. AI sector, AI advancements in China are progressing rapidly. Alibaba's Qwen Team, known for developing and releasing a growing family of powerful Qwen open-source language and multimodal AI models, has introduced its newest batch: the Qwen3.5 Small Model Series.
This series includes several models tailored for different applications:
- Qwen3.5-0.8B & 2B: These two models are optimized for "tiny" and "fast" performance, designed for prototyping and deployment on edge devices where battery life is a critical factor.
- Qwen3.5-4B: This model serves as a strong multimodal base for lightweight agents and natively supports a substantial 262,144 token context window.
- Qwen3.5-9B: A compact reasoning model that has shown to outperform OpenAI's open-source gpt-oss-120B, which is 13.5 times larger, on crucial third-party benchmarks. These benchmarks include multilingual knowledge and graduate-level reasoning capabilities.
To provide context, these models are among the smallest general-purpose models recently released by any lab globally. They are more comparable to MIT offshoot LiquidAI's LFM2 series, which also feature several hundred million or billion parameters, rather than the estimated trillion parameters reportedly used for flagship models from OpenAI, Anthropic, and Google's Gemini series.
The weights for these models are currently available worldwide under Apache 2.0 licenses on Hugging Face and ModelScope. This licensing makes them ideal for enterprise and commercial use, allowing for customization as needed.
The underlying technology of the Qwen3.5 small series represents a departure from standard Transformer architectures. Alibaba has adopted an Efficient Hybrid Architecture that integrates Gated Delta Networks (a form of linear attention) with sparse Mixture-of-Experts (MoE). This hybrid approach is designed to address the "memory wall" that typically restricts the performance of smaller models.