Diffusion-Qwen3.8-27B: Qwen3 adapted into a masked diffusion LM with bidirectional attention
GabGarrett · x · 2026-11-06
An independent researcher released Diffusion-Qwen3.8-27B, adapting Qwen3.8-27B into a masked diffusion language model: the prompt stays fixed, generation starts from fully masked tokens, and each pass predicts multiple missing tokens in parallel, committing only those above a confidence threshold as context for the next pass.
To enable bidirectional context, the causal mask was removed from full-attention layers, and Gated DeltaNet layers scan both directions with a learned gate weighting the reverse scan.
It's a research checkpoint with no accuracy/speed benchmarks yet. A noted failure mode: an early confident mistake can lock in errors and later passes build fluent but wrong output.
More from Models
- SGLang Ships Day-0 Kimi K3.1 Support With Out-of-the-Box Megakernel — BanghuaZ · 2026-11-06
- Google Launches Gemini Omni 1.2 Flash: 15-Second Video at Lower Pricing — FinanceYF5 · 2026-11-06
- Leaker DM claims Qwen4 is dropping, but even the timeline is a guessing game — bclavie · 2026-11-06
- All 22 DeepSeek V4.1 Flash endpoints statistically in line, most consistent model in the Index — ArtificialAnlys · 2026-11-06
- All 22 DeepSeek V4.1 Flash endpoints match reference — 3.9-point range, most consistent model in the Index — ArtificialAnlys · 2026-11-06
- Artificial Analysis launches Endpoint Accuracy Index v1.1: nearly 1 in 4 serverless endpoints score below reference — ArtificialAnlys · 2026-11-06