Skip to main content
Back

Alibaba's Qwen3.5-9B Beats OpenAI's 120B Model on Graduate Reasoning

Alibaba's Qwen team ships a 9B open-weight model that matches gpt-oss-120B on reasoning benchmarks while running on a laptop, natively multimodal.

Pro

Model, Full Access
  • Alibaba released Qwen3.5-9B under Apache 2.0, a compact reasoning-focused model in the Qwen3.5 Small Series.
  • Scores 81.7 on GPQA Diamond, beating OpenAI's gpt-oss-120B (80.1) at 13x smaller size.
  • Hybrid architecture combines Gated Delta Networks linear attention with sparse Mixture-of-Experts for efficiency.
  • Natively multimodal via early fusion on visual and textual tokens, no bolted-on CLIP encoder.
  • Trained with Scaled Reinforcement Learning across million-agent environments for structured reasoning.
  • Runs locally on laptops; both Base and Instruct versions live on Hugging Face.

Alibaba's Qwen team has quietly shipped one of the most aggressive efficiency plays in the open-weight race so far. The new Qwen3.5-9B lands as the flagship of a four-model small series, and its headline claim is uncomfortable for the trillion-parameter camp: on several third-party benchmarks it beats models more than ten times its size, while fitting on a MacBook.

The Qwen3.5 Small Model Series spans four tiers. Qwen3.5-0.8B and 2B target edge devices where battery life matters. Qwen3.5-4B is a multimodal base for lightweight agents with a 262,144 token context window. Qwen3.5-9B sits at the top as the reasoning-focused flagship, a compact model that outperforms the 13.5x larger OpenAI gpt-oss-120B on multilingual knowledge and graduate-level reasoning tests.

Small model, embarrassing benchmarks

The scores are the part that will make people re-read the model card. On visual reasoning, Qwen3.5-9B hit 70.1 on MMMU-Pro, ahead of Gemini 2.5 Flash-Lite at 59.7 and the specialized Qwen3-VL-30B-A3B at 63.0. Text reasoning is similarly loud:

  • GPQA Diamond: 81.7, surpassing gpt-oss-120b at 80.1
  • Video-MME with subtitles: 84.5, well clear of Gemini 2.5 Flash-Lite at 74.6
  • HMMT Feb 2025 mathematics tournament: 83.2
  • OmniDocBench v1.5 document recognition: 87.7
  • MMMLU multilingual: 81.2, versus gpt-oss-120b at 78.2

GPQA Diamond is the number to focus on. It is a set of graduate-level science questions written to be Google-proof. A 9B model beating a 120B one there is not a rounding error, it is a signal that parameter count has stopped being a reliable proxy for reasoning quality in this size class.

What is actually under the hood

The architecture is where the efficiency comes from, and it is no plain Transformer. Alibaba has moved toward an Efficient Hybrid Architecture that combines Gated Delta Networks, a form of linear attention, with sparse Mixture-of-Experts. Linear attention scales with sequence length instead of its square, so long contexts stop being a memory catastrophe. The MoE half means only a subset of parameters activates per token, keeping inference cheap.