Qwen's 4B Edge Model Tops Claude 4 in AIME25

Qwen's 4B edge AI models beat Claude 4 & Gemini 2.5 Pro in AIME25. Lightweight, powerful, runs on Raspberry Pi!

"AI Disruption" Publication 7300 Subscriptions 20% Discount Offer Link.

---

The Qwen team gets restless if they don't open-source for three days.

Last night, they released two more edge models:

• Qwen3-4B-Instruct-2507: A non-reasoning model with significantly enhanced general capabilities

• Qwen3-4B-Thinking-2507: An advanced reasoning model designed specifically for expert-level tasks

The 4B size is extremely edge-friendly, even meaning it can run on a Raspberry Pi.

In Qwen's official tweet, these two models are described as: Smarter, more precise, and supporting 256k context with enhanced context awareness.

On AIME25, Qwen3-4B-Thinking-2507 achieved remarkable results with minimal resources, scoring 81.3.

This score already surpasses Gemini 2.5 Pro (49.8~88.0) and Claude 4 Opus (75.5) on AIME25 — with just a 4B model!

Moreover, the 4B parameter count means it's extremely edge-friendly.