NVIDIA's Hugging Face org ships open Nemotron lineup: 30B/3B hybrid MoE, 1M context, edge-ready
mervenoyann · x · 2026-10-24
Hugging Face's merve poked fun at NVIDIA for not checking its own HF page, spotlighting NVIDIA's official Hugging Face org — the home for its open model weights, datasets, and interactive demos spanning generative AI, Physical AI, and agentic workflows.
Nemotron family highlights:
- Nemotron 3.5 Lightning: an efficient SLM with a hybrid Mamba-2 + Transformer MoE architecture (30B total / 3B active) plus Multi-Token Prediction layers
- Optimized for long-running autonomous agents and local inference on a single H100 or DGX Spark (GB10), with 1M-token context and reasoning ON/OFF modes
- Ships speculative-decoding drafters (DSpark, DFlash, MTP); servable via vLLM, SGLang, TRT-LLM, Ollama, llama.cpp
- All models come with open weights, open datasets, and reproducible training recipes across language, vision, retrieval, speech, and safety
More from Infra
- Google Cloud launches CLI Remote MCP Server for agents to run gcloud & bq without local installs — rseroter · 2026-10-24
- Baseten launches Environment Network, pairing vetted RL environment firms with post-training customers — baseten · 2026-10-24
- Bittensor-powered Engy hits 103B daily tokens, undercutting official API prices by up to 90% — markjeffrey · 2026-10-24
- DeepSeek serves 150T tokens/day on one cluster, more than OpenRouter, Fireworks, Together and BaseTen combined — natolambert · 2026-10-24
- Gradium open-sources Phonon: 31M-param on-device TTS with 0.02s latency, Apache 2.0 — natesiggard · 2026-10-24
- Meta and NVIDIA on Prometheus: gigawatt training cluster cut failures by 99% — BenBajarin · 2026-10-24