NVIDIA's research team published a July 2026 arXiv preprint on July 7 proving that autoregressive and diffusion language models can be trained together on the same objective, producing a single model that switches between three generation modes at inference time without any weight changes. The result, available as free open weights on Hugging Face under a commercial license, lets engineers running speculative decoding replace the separate draft model with a single checkpoint. The paper introduces Nemotron-Labs-Diffusion, a model family trained simultaneously on autoregressive and masked-diffusion objectives. The 8B instruct variant decodes six times more tokens per forward pass than a comparable AR model and hits four times higher throughput on NVIDIA's serving benchmark in self-speculation mode when deployed with the SGLang inference framework on a GB200 GPU — results from a preprint that has not yet undergone peer review or independent replication on other hardware. The model addresses the structural problem of low concurrency, where each token step leaves the GPU largely idle in the memory-bandwidth-bound regime. Starting from Mistral AI's Ministral3 base model, Nemotron-Labs-Diffusion's central claim is that autoregressive and diffusion paradigms are complementary.