256K
0.195
1.56
Chat
Active
Qwen3-Next-80B-A3B Thinking
It supports multi-token prediction and large context windows (up to 1 million tokens), enabling efficient real-time reasoning and interactive applications.
-p-130x130q80-p-130x130q80.png)
Qwen3-Next-80B-A3B Thinking integrates seamlessly into modern AI workflows with flexible deployment options including serverless, on-demand dedicated, and reserved monthly instances.
Qwen3-Next-80B-A3B Thinking is a cutting-edge reasoning-focused chat model designed for complex multi-step problem solving and chain-of-thought tasks. It outputs structured “thinking” traces by default and excels at tasks requiring deep analytical reasoning such as math proofs, code synthesis, logic, and agentic planning.
Qwen3-Next-80B-A3B Thinking is a sophisticated language model featuring 80 billion parameters, with only 3 billion actively engaged per token thanks to a sparse Mixture of Experts (MoE) architecture. It comprises 48 layers with a 2048-hidden dimension and employs a hybrid design combining gating mechanisms and advanced normalization techniques like RMSNorm. The model supports an expansive context window of 262K tokens, extensible up to 1 million tokens with specialized scaling methods, enabling superior long-context understanding. Trained with resource-efficient hybrid strategies, it achieves high performance in complex reasoning, math, coding, and multi-step problem solving, while maintaining low inference costs and high throughput, especially in tasks requiring deep analytical capabilities.
Input: $0.195
Output: $1.56
vs Qwen3-32B: Qwen3-Next-80B-A3B activates only 3 billion parameters per token compared to Qwen3-32B's full activation, making it about 10 times more efficient in training and inference cost. It also delivers over 10 times faster output speed in long-context scenarios (beyond 32K tokens) while achieving higher accuracy on reasoning and complex tasks.
vs Qwen3-235B: Despite having fewer active parameters, Qwen3-Next-80B-A3B approaches the performance levels of the much larger 235 billion parameter Qwen3-235B, especially in instruction following and long-context reasoning. It offers a favorable balance of compute efficiency and high model quality suitable for production use.
vs Google Gemini-2.5-Flash-Thinking: The Qwen3-Next-80B-A3B Thinking variant outperforms Google Gemini-2.5-Flash-Thinking in chain-of-thought reasoning and multi-turn instruction tasks while maintaining substantially lower operational costs due to sparse activation and multi-token prediction capabilities.
vs Llama 3.1-70B: Qwen3-Next-80B-A3B offers better long-range context understanding and reasoning stability at much larger context windows (up to 1 million tokens scalable) compared to Llama 3.1-70B's shorter native window. The sparse MoE architecture also gives it better efficiency at scale.
Qwen3-Next-80B-A3B Thinking is a cutting-edge reasoning-focused chat model designed for complex multi-step problem solving and chain-of-thought tasks. It outputs structured “thinking” traces by default and excels at tasks requiring deep analytical reasoning such as math proofs, code synthesis, logic, and agentic planning.
Qwen3-Next-80B-A3B Thinking is a sophisticated language model featuring 80 billion parameters, with only 3 billion actively engaged per token thanks to a sparse Mixture of Experts (MoE) architecture. It comprises 48 layers with a 2048-hidden dimension and employs a hybrid design combining gating mechanisms and advanced normalization techniques like RMSNorm. The model supports an expansive context window of 262K tokens, extensible up to 1 million tokens with specialized scaling methods, enabling superior long-context understanding. Trained with resource-efficient hybrid strategies, it achieves high performance in complex reasoning, math, coding, and multi-step problem solving, while maintaining low inference costs and high throughput, especially in tasks requiring deep analytical capabilities.
Input: $0.195
Output: $1.56
vs Qwen3-32B: Qwen3-Next-80B-A3B activates only 3 billion parameters per token compared to Qwen3-32B's full activation, making it about 10 times more efficient in training and inference cost. It also delivers over 10 times faster output speed in long-context scenarios (beyond 32K tokens) while achieving higher accuracy on reasoning and complex tasks.
vs Qwen3-235B: Despite having fewer active parameters, Qwen3-Next-80B-A3B approaches the performance levels of the much larger 235 billion parameter Qwen3-235B, especially in instruction following and long-context reasoning. It offers a favorable balance of compute efficiency and high model quality suitable for production use.
vs Google Gemini-2.5-Flash-Thinking: The Qwen3-Next-80B-A3B Thinking variant outperforms Google Gemini-2.5-Flash-Thinking in chain-of-thought reasoning and multi-turn instruction tasks while maintaining substantially lower operational costs due to sparse activation and multi-token prediction capabilities.
vs Llama 3.1-70B: Qwen3-Next-80B-A3B offers better long-range context understanding and reasoning stability at much larger context windows (up to 1 million tokens scalable) compared to Llama 3.1-70B's shorter native window. The sparse MoE architecture also gives it better efficiency at scale.
.webp)
.webp)
.png){kind=small}
.png){kind=small}
