Qwen 3.5 35B A3B

• 🧠 35B-parameter mixture-of-experts activating only ~3B per token
• 📏 Native 256K-token context window
• 👁️ Multimodal model handling both text and vision inputs
• 🔧 Built for reasoning, coding, agents, function calling, web search
• 🏢 From Alibaba's Qwen team, released under Apache 2.0
• 🆕 Provider reports it surpasses the larger Qwen3-235B-A22B
• ⚡ Sparse activation targets efficient, lower-cost inference

Alibaba

Alibaba Group is a Chinese multinational technology company founded in 1999 and headquartered in Hangzhou, Zhejiang. Originally built around e-commerce and cloud computing, Alibaba has become one of the most prolific contributors to open-weight AI research,…

Qwen 3.5 35B A3B is a sparse mixture-of-experts model from Alibaba's Qwen team, released in February 2026 under the Apache 2.0 license. It carries roughly 35 billion total parameters but activates only about 3 billion per token, a design intended to deliver large-model behavior at a fraction of the compute cost. The model is multimodal, accepting text and vision inputs, and supports a native 256K-token context window alongside reasoning, code-optimized generation, function calling, and web search.

Within the Qwen family, this release sits between smaller and larger 3.5-generation siblings such as Qwen 3.5 9B and the much larger Qwen 3.5 397B. The provider describes the 35B A3B model as surpassing the earlier, denser Qwen 3 235B A22B Instruct 2507 while being roughly 6.7 times smaller in total parameters — a generational efficiency gain attributed to the company's own description.

In practice, the model targets reasoning, coding, and general-knowledge tasks where its small active-parameter footprint keeps latency and serving costs low. Later Qwen releases such as Qwen 3.6 27B continue this compact mixture-of-experts direction. Because primary benchmark documentation specific to this checkpoint is limited, the description here stays to verifiable architectural and licensing facts.