nvidia and NVIDIA Introduce Nemotron Labs-3 Puzzle with 75B A9B through NVFP4 on Hugging Face
huggingface inference nvidia
| Source: Mastodon | Original article
NVIDIA releases Nemotron-Labs-3-Puzzle model for optimization. It features LatentMoE and NVFP4 quantization for improved serving.
NVIDIA has published Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4, a model designed for serving optimization. This development is significant as it utilizes LatentMoE with Mamba-Interleaving, Multi-Token Prediction, and NVFP4 quantization to enhance performance. The OpenMDW-1.1 license allows for commercial use, and the model has achieved an AIME25 score of 89.9.
This release matters because it demonstrates NVIDIA's ongoing efforts to improve the efficiency and accuracy of its AI models. The Nemotron family of models is known for its open-source nature, making it accessible for developers to build specialized AI agents with reasoning capabilities. As we reported on July 5, the community has been actively exploring various hardware configurations for running AI models, and this new development may have implications for those efforts.
As the AI landscape continues to evolve, it will be interesting to watch how NVIDIA's Nemotron models are adopted and utilized by developers. With the release of Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4, we can expect to see further innovations in serving optimization and interactive deployment. The ability to compress hybrid MoE LLMs, as seen in this model, may also lead to new breakthroughs in AI research and applications.
Sources
Back to AIPULSEN