LLaDA2.2-mini

Full-precision diffusion with native token rewriting, insertion and deletion.

Prompt

Ready

System prompt

New tokens

1 512

Denoising steps per block

Controls the minimum unmasking schedule. Refinement may add more iterations.

1 128

Block length

Must align to the model's 32-token MoE routing blocks.

Temperature

0 is deterministic. Native mode samples masked tokens; BG + TG adds Gumbel noise to ordinary-token paths.

0 2

Bigram weight

0 5

Top-k

0 means all tokens in native mode; BG + TG uses 16 candidates when set to 0.

0 100

Token coloring

No coloring Prediction probability Prediction iteration

Decoding method

Confidence-guided BG + TG Viterbi (Gumbel)

Early stopping

Stop after a completed block contains EOS. Blocks may converge sooner when stable.

INSERT opens a mask before the previous token; DELETE shifts tokens left. Blocks retain their size by truncating or padding with masks. BG + TG guides ordinary text; structural edit predictions remain native.

Model card · Source code