Alibaba's Qwen team just released the Qwen3.5 Medium series, three open-source models that pick a fight with one of the AI industry's core assumptions: that frontier-level performance requires a cloud data center. The flagship runs on a single consumer GPU with 32GB of VRAM, handles more than a million tokens of context, and still beats models several times its size on key benchmarks.
All three models ship under the permissive Apache 2.0 license, which means anyone — a startup, an agency, a solo developer — can download, run, fine-tune, and deploy them commercially without asking permission. That combination of local-first performance and open licensing is the real story here, and it quietly reshapes where AI answers about your brand will come from.
Key takeaways
- Qwen3.5-35B-A3B, the flagship, exceeds 1M tokens of context on a consumer GPU (32GB VRAM), enabled by near-lossless 4-bit quantization — capabilities that previously demanded server-grade hardware. - On benchmarks it outperforms GPT-5-mini across multiple tests, surpasses Claude Sonnet 4.5 in knowledge (MMMLU) and visual reasoning (MMMU-Pro), and beats Qwen's own 235B predecessor with a fraction of the active parameters. - The lesson is architectural: a hybrid of Gated DeltaNet and Mixture-of-Experts (MoE) shows that efficient design can beat raw scale. - For GEO: cheaper, local, open models multiply the number of AI surfaces that answer questions about your brand. Visibility is no longer a ChatGPT-and-Gemini problem — it fragments across a widening field of engines and markets, and it has to be tracked that way.
What Alibaba shipped
The Qwen3.5 Medium series is three models, each tuned for a different deployment envelope.
Qwen3.5-35B-A3B is the flagship. It carries 35B total parameters but activates only 3B per token, supports 1M+ tokens of context, and is built to run on consumer GPUs with 32GB of VRAM.
Qwen3.5-122B-A10B is the heavyweight: 122B total, 10B active, 1M+ token context, aimed at server-grade GPUs with 80GB of VRAM.
Qwen3.5-27B is the efficiency play: a dense 27B model with 800K+ token context, built for high-throughput, low-overhead deployment.
The headline number is the flagship's context length. Processing more than a million tokens has historically meant renting expensive server hardware. Qwen3.5-35B-A3B does it on a GPU you could put in a workstation, thanks to near-lossless 4-bit quantization that shrinks the memory footprint without gutting quality.
Beating bigger models
Qwen didn't just match the field; on several axes it moved ahead of it. The 35B-A3B model outperforms GPT-5-mini across multiple benchmarks. It surpasses Claude Sonnet 4.5 specifically in knowledge, measured by MMMLU, and in visual reasoning, measured by MMMU-Pro. Most telling, it beats Qwen's own previous-generation 235B model while activating a tiny fraction of the parameters.
