[ICLR 2026] Official repository of "Beyond Fixed: Training-Free Variable-Length Denoising for Diffusion Large Language Models"
A 4B diffusion grounding model for fast, precise visual evidence extraction. Unifies diverse grounding tasks through parallel decoding, with optional self-speculative decoding delivering 4.51x autoregressive throughput on COCO.
[AAAI 2026 & ACL 2026] The official implementation of the DIFFA series for dLLM-based large audio language model
Embedding Inversion via Conditional Masked Diffusion: recover original text from embedding vectors using parallel denoising. Live demo + training pipeline + technical report.
[ICLR2026] Official Code for Loopholing Discrete Diffusion: Deterministic Bypass of the Sampling Wall
A curated list of diffusion language models (dLLMs): papers, models, and code behind diffusion-based, non-autoregressive text generation, focused on 2025–2026.
Predictive-State Flows (PSF): a flow prior generates K predictive states, each summarizing only the preceding blocks; a block decoder writes all blocks in parallel. Code and checkpoints for arXiv 2610.12898.
[EMNLP 2026] Length-Adaptive Decoding for Masked Diffusion Machine Translation — a training-free target-length selector for masked diffusion LMs.
A compact, from-scratch trainer for masked diffusion and GPT-style autoregressive language models in one shared PyTorch Transformer.
🛠 Reconstruct original text from text embeddings using conditional masked diffusion to reveal reversible embedding representations efficiently and accurately
DiffusionGemma node pack for ComfyUI, built on MCP backbone for fully-agentic consumption — discrete diffusion text generation with per-step canvas snapshots, commit heatmaps, and structured trace data. Watch meaning crystallize out of noise.
Jev-style typed decision engine on a diffusion LM (DiffusionGemma + MLX): one-pass slot reads, calibration, domain LoRA. Not affiliated with TypeSafe AI.
Local, type-aware diffusion code editor — on-device whole-region code edits steered by type-checker feedback inside the denoising loop (Apple Silicon / MLX)
Minimal masked diffusion LM in PyTorch. Pretraining + SFT + a scaling family from 50M to 350M, LLaDA recipe.
Run NVIDIA's Nemotron-Labs-TwoTower-30B-A3B block-wise diffusion LM natively on Apple Silicon with MLX — AR tower + full two-tower mask-diffusion, 4/6/8-bit + bf16.
Measured llama.cpp and quantization results on Tesla V100 (sm_70), Pascal GTX 1070, and RTX 4070 - hardware most projects don't test on
Clover-style diffusion language model (dLLM) reference implementation: discrete mask-and-predict diffusion, sparse MoE (top-k + load balance), confidence remasking, joint-mode decision. Pure Python, zero dependencies.
AR → discrete diffusion LLM backend (LLaDA 2.1 recipe, Qwen3-0.6B, Pascal/sm_61)
Clover-35B-A3B-Preview (block-diffusion MoE, 3B active, BF16) on a single DGX Spark with SGLang via Docker Compose. Includes an agent benchmark against Qwen3.8 Flash Next: BFCL, AgentDojo, IFEval, GPQA.
Diffusion LM + frozen AR context tower: can block-parallel diffusion reuse a context cache?
To associate your repository with the diffusion-language-model topic, visit your repo's landing page and select "manage topics."