Model Releases
Qwen3-Next: The 80B MoE Revolution for Local Deployment
Alibaba Cloud unleashes Qwen3-Next, an ultra-efficient 80B MoE model with only 3B active parameters, now available under Apache 2.0.
September 10, 2025
Model ReleaseQwen3-Next
Introduction
Alibaba Cloud has officially announced the release of Qwen3-Next, a groundbreaking open-source large language model that redefines efficiency in the current AI landscape. Released on September 10, 2025, this model targets developers seeking high-performance reasoning without the heavy computational overhead of standard dense transformers. By leveraging a sophisticated Mixture of Experts (MoE) architecture, Qwen3-Next delivers capabilities comparable to proprietary Sonnet 4.5 models while remaining accessible for local deployment. This release marks a significant shift in the open-source community, prioritizing parameter efficiency without sacrificing intelligence. The strategic move to open-source this architecture under the Apache 2.0 license invites widespread adoption and community contribution, potentially accelerating the development of specialized AI applications globally.
- Release Date: September 10, 2025
- License: Apache 2.0
- Provider: Alibaba Cloud
Key Features & Architecture
The architecture of Qwen3-Next is built on an 80 billion parameter backbone, but it only activates 3 billion parameters during inference. This MoE design allows the model to dynamically select the most relevant experts for specific tasks, reducing latency and energy consumption significantly. The model supports a massive context window of 128,000 tokens, enabling long-form document analysis and complex multi-step reasoning. Additionally, it retains full multimodal capabilities, allowing for seamless integration of text and image inputs in a unified pipeline. This dual-mode operation ensures that developers can handle both unstructured data and complex visual queries without needing separate models.
- Total Parameters: 80B
- Active Parameters: 3B
- Context Window: 128k tokens
- Architecture: MoE (Mixture of Experts)
Performance & Benchmarks
Benchmark results indicate a substantial leap in reasoning and coding capabilities compared to previous iterations. On the MMLU benchmark, Qwen3-Next scores 88.5, outperforming previous Qwen3.5-Medium variants and approaching top-tier proprietary models. In HumanEval, it achieves a pass rate of 92.3%, demonstrating robust code generation skills across multiple programming languages. Furthermore, on the SWE-bench repository, the model solves 45% of hard issues, proving its utility for software engineering tasks. These metrics suggest that the 3B active parameter count does not compromise the quality of output, maintaining high fidelity in complex logical tasks while consuming significantly less GPU memory.