Alibaba Unveils Qwen 3.5 Language Model Series
Alibaba's Qwen project announced the release of its Qwen 3.5 series of large language models, including Flash, 35B-A3B, 122B-A10B, and 27B variants. The company claims the models achieve superior intelligence through architectural improvements, data enhancements, and reinforcement learning, rather than sheer scale. The new series supports a one million-token context window and advanced tool usage.
- The Qwen project, which began in April 2023 under the name Tongyi Qianwen, initially based its architecture on Meta AI's Llama model. - Predecessor models like Qwen2.5 were pretrained on datasets of up to 18 trillion tokens, with the Qwen2.5-Max version trained on over 20 trillion tokens. - In performance benchmarks, the Qwen2.5-72B-Instruct model surpassed models like LLama 3.1 405B and Mistral Large 2 in areas such as coding and math problem-solving. - The Qwen family includes specialized models beyond general language tasks, such as Qwen2.5-Coder for programming and Qwen2.5-Math for mathematical reasoning. - The project has expanded into multimodality with models like Qwen2.5-VL, which can analyze images and videos, and Qwen2.5-Omni, which processes text, image, audio, and video inputs to generate text and speech responses. - Alibaba has pursued an open-source strategy for many Qwen models, leading to Qwen becoming the most downloaded AI model on the Hugging Face platform as of January 2026, surpassing Meta's Llama. - The Qwen3.5 series is positioned as more cost-effective, with Alibaba stating it is 60% cheaper and runs workloads eight times faster than the previous version. - For API usage, the smaller Qwen2.5 7B Instruct model is priced significantly lower than competitors like GPT-4, costing 100 times less for input processing.
