NVIDIA Unveils Nemotron-Labs-Diffusion

Nemotron-Labs-Diffusion supports AR decoding, diffusion-based parallel decoding, and self-speculation all within the same model, simply by changing the attention pattern.

The key idea is to shift the bottleneck of the generation stage from memory-bound to compute-bound, reusing weights loaded once across computation for multiple tokens.

The main points are as follows.

• Provided as a 3B, 8B, 14B dense LM family.
• Includes base, instruct, and vision-language variants.
• Self-speculation works by drafting with diffusion and verifying with AR.
• Claims 3x acceptance length and 2.2x speedup compared to Qwen3-8B-Eagle3 on SGLang.
• States it achieved 5.9x tokens per forward at the same accuracy.
• For measured performance, presents 112 tok/s on DGX Spark and 850 tok/s on GB200, explaining that custom CUDA kernels pushed this up to 1015 tok/s.

Additionally, through a diffusion speedup-of-light analysis, it suggests there is room for sampling improvements alone to further boost single-user throughput by 2x beyond the current best.