QWEN / STREAMING ASR / 2026

Qwen3-ASR-1.7B

TOTAL2.35Bparameters

ACTIVE2.35Bruntime path

DEPTH24 + 28repeated blocks

HIDDEN1,024→2,048main trunk

CONTEXT65Kstream / sequence

CHECKPOINT4.70GB2 shards · Apache-2.0

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

30 seconds → 256 tokens · 2.24T arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT30SECONDS

TRANSFORMER STACK24 + 28 LAYERSaudio encoder → projector → causal text decoder

Dual-Stack Residual

OUTPUT256TOKENS

wraps every repeated block

WINDOWED AUDIO ATTENTION · UNIFIED STREAMING同一模型通过不同窗口策略覆盖低延迟流式与长音频离线转写。

Qwen3 Decoder自回归解码器生成语言、文本与时间相关 token。

AUDIO24L1,024 hidden

TEXT28L2,048 hidden

HEADS16 / 8Q / KV

MEL128bins

CONTEXT65Ktokens

LANG52languages