Qwen Upgrades to 3B, 256K, Nears GPT-4o Performance
Qwen3-30B-A3B: 3B activated, 256K context, rivals GPT-4o & Gemini 2.5. Faster, stronger, open-source AI!
"AI Disruption" Publication 7200 Subscriptions 20% Discount Offer Link.
Happy QwensWeek truly lives up to its name.
Qwen's base models have been open-sourced one after another, and now the brand new non-thinking model Qwen3-30B-A3B-Instruct-2507 has also launched lightning-fast.
With only 3B parameters activated, it achieves exceptional performance comparable to top-tier closed-source models like Gemini 2.5-Flash (non-thinking) and GPT-4o.
Compared to the previous generation non-thinking model Qwen3-30B-A3B Non-Thinking, this "minor update" has brought critical improvements to the model's general capabilities.
Among these, the model's reasoning ability (AIME25) improved by 183.8%, while its alignment capability (Arena-Hard v2) improved by 178.2%. Additionally, the model's long-text processing capability has been upgraded from the previous generation's 128K to 256K.