QWEN / DENSE + MOE FAMILY / 2025

Qwen3-235B-A22B

TOTAL235Bparameters

ACTIVE22Bper token

DEPTH94repeated blocks

HIDDEN4,096main trunk

CONTEXT40K / 131Knative tokens

CHECKPOINT≈470GB94 shards · Apache-2.0

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

100K INPUT + 1K OUTPUT · 20.15P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT100.0KTEXT

TRANSFORMER STACK94 MOE BLOCKSGQA + Qwen MoE in every decoder layer

RMSNorm Residual

OUTPUT1.0KTOKENS

wraps every repeated block

GROUPED-QUERY ATTENTION · THINKING MODE SWITCH推理模式主要由后训练与模板控制，不是另一个主干。

Qwen MoE · 235B CAPACITY22B active · total capacity ≠ active compute

DEPTH94decoder

HIDDEN4,096main trunk

HEADS64 / 4Q / KV

EXPERTS128top-8

CONTEXT40Knative

PARAMS235Bflagship