main
official Qwen3-Next repository and 80B-A3B model config
QWEN / ULTRA-SPARSE HYBRID / 2025

Qwen3-Next-80B-A3B

TOTAL80Bparameters
ACTIVE3Bper token
DEPTH48repeated blocks
HIDDEN2,048main trunk
CONTEXT262Knative tokens
CHECKPOINT≈160GB32 shards · Apache-2.0
INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

Qwen3-Next-80B-A3B · Architecture Explorer

100K INPUT + 1K OUTPUT · 1.11P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER
INPUT100.0KTEXT
TRANSFORMER STACK48 HYBRID BLOCKS36 DeltaNet · 12 full GQA · ultra-sparse MoE
RMSNorm Residual
++
OUTPUT1.0KTOKENS
wraps every repeated block
GATED DELTANET · PERIODIC GLOBAL READ线性层与全局层不是并行分支，而是明确的深度调度。
Ultra-Sparse MoE · 80B CAPACITY≈3B active · total capacity ≠ active compute
+
DEPTH4836 + 12
HIDDEN2,048main trunk
SCHEDULE3:1linear/global
EXPERTS512top-10
CONTEXT262Ktokens
ACTIVE3Bper token