NVIDIA Nemotron-Labs-Diffusion-14B supports three decoding modes
The 14B model switches attention patterns to support autoregressive, diffusion, and linear self-speculative decoding. Diffusion was 2.7× faster than autoregressive decoding on DGX Spark and 3.3× on GB200. Self-speculation achieved 3× the acceptance length and 2.2× the SGLang throughput of Qwen3-8B-Eagle3. The family includes 3B, 8B, 14B, and vision-language variants under the NVIDIA Nemotron Open Model License.
Visit original source