Qwen3 Next 80B A3B

System Requirements

VRAM requirements for different quantization methods and context sizes

About Qwen3 Next 80B A3B

Qwen3-Next-80B-A3B is a high-capacity sparse Mixture-of-Experts (MoE) foundation model developed by Alibaba's Qwen team. It belongs to the next-generation Qwen3-Next series, specifically designed to address the computational demands of long-context sequence modeling and large-scale parameter efficiency. The model features a unique hybrid attention mechanism that integrates Gated DeltaNet with Gated Attention, allowing the system to maintain high performance across extended token sequences while significantly reducing the quadratic complexity typically associated with standard Transformer architectures.

The technical architecture employs a high-sparsity MoE layout consisting of 48 layers with a hidden dimension of 2048. While the model contains 80 billion total parameters, its gating mechanism activates only approximately 3 billion parameters per token during inference. This sparse activation strategy, combined with a total of 512 experts and a multi-token prediction (MTP) objective, facilitates improved throughput and reduced FLOPs per token. The model also incorporates stability-focused architectural refinements, such as zero-centered and weight-decayed layer normalization, to ensure robust convergence during both pre-training on 15 trillion tokens and subsequent reinforcement learning stages.

Optimized for complex reasoning and agentic workflows, Qwen3-Next-80B-A3B is capable of processing a native context window of 262,144 tokens, which can be extended to over 1 million tokens using specialized scaling techniques like YaRN. Its primary use cases include multi-step logical analysis, mathematical proofs, and code synthesis. By separating the 'Thinking' variant, which outputs structured reasoning traces, from the standard 'Instruct' variant, the model provides specialized paths for either high-efficiency general-purpose interaction or intensive, transparent problem-solving tasks.

Technical Specifications

Resources

About Qwen 3

The Alibaba Qwen 3 model family comprises dense and Mixture-of-Experts (MoE) architectures, with parameter counts from 0.6B to 235B. Key innovations include a hybrid reasoning system, offering 'thinking' and 'non-thinking' modes for adaptive processing, and support for extensive context windows, enhancing efficiency and scalability.

Other Qwen 3 Models

• Qwen3-0.6B
• Qwen3-1.7B
• Qwen3-4B
• Qwen3-8B
• Qwen3-14B
• Qwen3-30B-A3B
• Qwen3-32B
• Qwen3 235B A22B Thinking
• Qwen3-235B-A22B
• Qwen3 Coder 480B A35B