NVIDIA AI Releases Star Elastic: One Checkpoint that Contains 30B, 23B, and 12B Reasoning Models with Zero-Shot Slicing

Short summary

NVIDIA's Star Elastic packs 30B, 23B, and 12B parameter reasoning models into a single checkpoint via post-training, reducing training tokens by 360× versus separate training. Elastic budget control uses a smaller submodel for reasoning and the full model for final answers, achieving 16% higher accuracy and 1.9× lower latency. FP8 and NVFP4 quantization bring the entire model family to consumer RTX GPUs.

• •Single checkpoint contains three nested model sizes (30B, 23B, 12B) trained efficiently in one run
• •Elastic budget control improves accuracy 16% and reduces latency 1.9× compared to standard approaches
• •Full model family becomes accessible on consumer RTX-class GPUs through quantization

Generated with AI, which can make mistakes.