Qwen3-Next
Qwen3-Next explores a high-efficiency hybrid-attention MoE architecture using Gated DeltaNet and gated attention. The 80B model activates 3B parameters per token, with separate Instruct and Thinking checkpoints and a native 256K context. Multi-token prediction and a high-sparsity expert design improve inference efficiency. Open Apache-2.0 weights and reference serving instructions support long-context research and deployment. The non-thinking Instruct and reasoning-focused Thinking variants have different chat behavior, which callers should handle explicitly.
2025-09-11
80B total, 3B activated
Mixture of Experts (MoE) with hybrid Gated DeltaNet and gated attention
Apache-2.0
Specifications
- Parameters
- 80B total, 3B activated
- Architecture
- Mixture of Experts (MoE) with hybrid Gated DeltaNet and gated attention
- License
- Apache-2.0
- Context Window
- 262,144 tokens
- Type
- text
- Modalities
- text
Benchmark Scores
Advanced Specifications
- Model Family
- Qwen
- API Access
- Not Available
- Chat Interface
- Not Available
- Variants
- Qwen3-Next-80B-A3B-InstructQwen3-Next-80B-A3B-Thinking
Capabilities & Limitations
- Capabilities
- reasoningcode generationtool usehybrid linear attentionlong-context modeling
- Known Limitations
- Self-hosting requires sufficient memory for all model weights, including inactive experts.Reasoning and agent benchmark results depend on the harness and sampling settings.Instruct is non-thinking; Thinking is a separate checkpoint.Context beyond the native window requires YaRN configuration.
- Notable Use Cases
- coding assistantsresearch agentsdocument analysisefficient inference research
- Tool Use Support
- Yes