Qwen3-Next-80B-A3B

Qwen3-Next-80B-A3B is Alibaba's late-2025 testbed for hybrid linear attention, released September 11, 2025 in Base, Instruct, and Thinking variants. It pairs 80 billion total parameters with only 3 billion active per token across 512 fine-grained experts, and replaces most vanilla attention layers with Gated DeltaNet, keeping gated attention only where recall matters. Multi-token prediction speeds up generation, and the native context window is 262,144 tokens. The team reports roughly 10x faster inference than Qwen3-32B on long contexts and reasoning that rivals the much larger Qwen3-235B. It ships under Apache 2.0. As a research-architecture release it is less battle-tested in production than the main Qwen3 line, and the hybrid attention design means some inference stacks need updates to run it well.

Team-reported: beats Qwen3-32B on downstream tasks; Thinking variant outperforms Gemini-2.5-Flash-Thinking (team-reported)