BREAKING
Atomic Chat Adds DFlash Decoding
Qwen3.6-27B on RTX 6000
Baseline
44
MTP
65
DFlash
98
0
x
average gain
0
tok/s
peak on JSON
0
x
peak speedup
How DFlash Works
1
Block-diffusion drafter
↓
2
Propose 15 tokens
↓
3
Target verifies
Strong for Code, Weak for Prose
Works well
structured
●
Coding tasks
●
JSON output
●
Short context
Unreliable
creative
●
Long-form stories
●
Low acceptance
●
Large context windows
Speculative Decoding Goes Local
AI NEWS BLITZ
Atomic Chat has added DFlash speculative decoding, claiming much faster local inference.