official model card + Diffusion-Gemma ASR architecture article
INTERFAZE / DIFFUSION ASR / 2026
Diffusion-Gemma-ASR Small
TOTAL26B + 42Mparameters
ACTIVE≈4Bruntime path
DEPTHWhisper + 48repeated blocks
HIDDEN768→2,816main trunk
CONTEXT256 canvasstream / sequence
CHECKPOINT191.08MB adapter1 shards · Apache-2.0
INPUT30SECONDS
TRANSFORMER STACKFROZEN ENCODER + 48L MOE42M projector/adapter is the only trainable checkpoint
Diffusion Residual
++
OUTPUT256TOKENS
wraps every repeated block
DIFFUSIONGEMMA MOE · PARALLEL TOKEN DENOISE所有文本位置并行更新,延迟随 pass 数而不是输出长度线性增长。
Audio Projector · 26B FROZEN CAPACITY≈4B active · total capacity ≠ active compute
BACKBONE26B-A4Bfrozen
ADAPTER42Mtrainable
CANVAS256tokens
PASSES8–16denoise
AUDIO30schunk
PROJECT8×subsample