NVIDIA Nemotron-Labs-Sigma generates token blocks in parallel
The non-obvious shift is that diffusion-style continuous generation is moving from images into open language models, breaking the one-token-at-a-time assumption that shapes most inference stacks. Developers building math or coding agents should benchmark the 3B/8B checkpoints against autoregressive alternatives like Qwen or DeepSeek, since parallel block generation trades sequential decoding for throughput. Open weights on Hugging Face make direct evaluation immediate.