> ## Content Index
> Fetch the complete content index at: https://www.theleftshift.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Alibaba's New AI Model Delivers 10x Efficiency and Outperforms Gemini on Reasoning
- URL: https://www.theleftshift.com/alibabas-new-model-delivers-10x-efficiency-and-outperforms-gemini-on-reasoning/
- Published: 2025-09-12T06:12:11.000Z
- Updated: 2025-09-12T06:49:09.000Z
- Description: The company says the new design brings dramatic gains in both efficiency and performance.
- Author: The Left Shift Bureau
- Tags: AI Models

Chinese e-commerce giant Alibaba has announced **Qwen3-Next-80B-A3B-Base** model with a [new AI architecture](https://qwen.ai/blog?id=4074cca80393150c248e508aa62983f9cb7d27cd&from=research.latest-advancements-list&ref=theleftshift.com) designed to tackle two major challenges shaping the future of large models — **Context Length Scaling** and **Total Parameter Scaling**. 

The company says the new design brings dramatic gains in both efficiency and performance, particularly for ultra-long context tasks and large-scale reasoning.

Unlike the Mixture-of-Experts (MoE) approach used in Qwen3, Qwen3-Next introduces a **hybrid attention mechanism**, a highly sparse MoE structure, multi-token prediction for faster inference, and optimisations for training stability. 

> 🚀 Introducing Qwen3-Next-80B-A3B — the FUTURE of efficient LLMs is here!  
>  
> 🔹 80B params, but only 3B activated per token → 10x cheaper training, 10x faster inference than Qwen3-32B.(esp. @ 32K+ context!)  
> 🔹Hybrid Architecture: Gated DeltaNet + Gated Attention → best of speed &… [pic.twitter.com/yO7ug721U6](https://t.co/yO7ug721U6?ref=theleftshift.com)
> 
> — Qwen (@Alibaba\_Qwen) [September 11, 2025](https://twitter.com/Alibaba%5FQwen/status/1966197643904000262?ref%5Fsrc=twsrc%5Etfw&ref=theleftshift.com)

These innovations allow the new **Qwen3-Next-80B-A3B-Base** model to activate only **3 billion parameters at inference**, while still matching or exceeding the performance of the dense **Qwen3-32B** model — and at less than **10% of its training cost**. For long context lengths above 32K tokens, the model achieves **10x higher throughput**, making it significantly more efficient for both training and deployment.

Building on this foundation, Qwen has released two post-trained variants:

- **Qwen3-Next-80B-A3B-Instruct**, which delivers performance on par with the flagship **Qwen3-235B-A22B-Instruct-2507**, with clear advantages for tasks requiring up to **256K tokens** of context.
- **Qwen3-Next-80B-A3B-Thinking**, which excels at complex reasoning tasks, outperforming lower-cost models like **Qwen3-30B** and even surpassing **Google’s Gemini-2.5-Flash-Thinking** on multiple benchmarks, while approaching the capabilities of Qwen’s top-tier **235B-Thinking** model.

Qwen highlights that the new architecture also solves **long-standing efficiency and stability issues** in reinforcement learning training, accelerating RL speed and improving final performance.

*"Qwen3-Next represents our latest exploration in hybrid model. By combining Gated DeltaNet and standard attention in a 3:1 ratio, we achieve stronger in-context learning and better overall performance," Binyuan Hui, AI researcher at Alibaba, said.*

Earlier this month, [Alibaba released its most powerful large language model yet](https://www.theleftshift.com/alibaba-unveils-1-trillion-parameter-qwen3-max-preview-model/), Qwen3-Max-Preview (Instruct), boasting more than 1 trillion parameters. In July, the company also [officially launched Qwen3-Coder](https://www.theleftshift.com/alibaba-enters-the-vibe-coding-scene-with-qwen-3-coder/), its most advanced open-source coding model.