Instead of generating every position in a block strictly one token at a time, Clover starts with masked tokens and resolves confident positions in parallel. Its adapted Qwen3.6-derived hybrid backbone supports bidirectional attention within the current block while retaining a causal prefix for completed text.
Clover is useful for developers studying diffusion-based language generation and alternative inference tradeoffs. The release supplies custom Transformers code and backend implementations, exposing block size, confidence-based decoding, and supported acceleration paths for experimentation with self-hosted language applications.
