The model is a 74-billion-parameter (74B) Mixture-of-Experts (MoE) model with 9 billion (9B) active parameters. It is compressed down to 24GB via 2-bit quantization using advanced quantization-aware training techniques to preserve model quality. Designed to strike the optimal balance between speed and intelligence, it runs faster than most mainstream dense models at a similar memory footprint while being smarter than most mainstream MoE models.
Nemotron 3.5 Bolt E4B 4GB will be released on October 23
Additional Details
Top Comments
beneath_steel_sky (⬆69): llama.cpp support https://github.com/ggml-org/llama.cpp/pull/30520 jacek2023 (⬆31): Day 0 support incoming Harley-Riders (⬆27): They really should support deepseek-ai/DeepSeek-V4.1-Flash by now—it’s been a whole month already. Why does new model support always land without MTP (Multi-Token Prediction)? Time to switch tools. Ok_Mammoth589 (⬆27): Hugging Face is just a poor startup; you can’t realistically expect their LLM-serving product to actually run the models. Embarrassed_Adagio28 (⬆144): Given that NVIDIA hasn’t dropped a single genuinely good model so far, I wouldn’t get my hopes up. JoeyDee86 (⬆102): That’s because almost nobody bothers to read what NVIDIA intends these models for. They are built as a “canvas”—the whole point is LoRA, allowing anyone with an NVIDIA GPU to fine-tune them on their own. Constant-Simple-1234 (⬆65): Exactly, and NVIDIA’s stance on open weights is just as respectable. Thrumpwart (⬆24): Correct me if I’m wrong, but aren’t Nemotron models specifically marketed as base models intended for users to fine-tune for their specific use cases?