NVIDIA Labs tri-mode LLM family: at each step the model can run as autoregressive, diffusion, or self-speculation, with a learned controller picking the regime. Reports 2.6–6.4× throughput over AR baselines while matching or beating Qwen3 accuracy on standard benchmarks.

Three sizes (3B / 8B / 14B), each with Base, Instruct, and VL variants; eight model uploads total on May 23. Released under the Nemotron Open Model License.

NLD-Uni (paper October 27, 2026, with weights on HuggingFace and code at NVlabs) extends the line to a unified discrete diffusion LLM for both visual understanding and generation. Two changes carry it. A cross-paradigm on-policy distillation transfers understanding from autoregressive VLMs and image generation from continuous flow models into one diffusion LLM. A progressive self-refining diffusion process corrupts each token over several steps instead of jumping straight from a clean image token to a mask, which lets the model commit to a token at an intermediate step and refine it later. NVIDIA reports state-of-the-art results among unified diffusion LLMs on VQA, multimodal reasoning, grounding, text-to-image and image editing.

Model Details

Parameters 14B

Variants

Name Parameters Notes
Nemotron-Labs-Diffusion-3B (Base/Instruct/VL) 3B —
Nemotron-Labs-Diffusion-8B (Base/Instruct/VL) 8B —
Nemotron-Labs-Diffusion-14B (Base/Instruct) 14B —
foundationaldiffusionopen-weightefficiency

Related