NVIDIA pretrained 120B and 550B MoEs in NVFP4, keeping last 15% and embeddings in BF16
Nouamanetazi · x · 2026-11-06
The MoE handbook examines 4-bit pretraining at scale: NVIDIA trained Nemotron 3 Super (120B total, 12B active) on 25T tokens and Nemotron 3 Ultra (550B, 55B active) on 20T tokens in NVFP4, keeping the last 15% of the network, attention projections and embeddings in BF16. An accompanying figure details DeepSeek's FP8 split.
Related event: Hugging Face Releases Free MoE Training Handbook Built on 800K GPU Hours(20 posts)→
More from Infra
- Qwen3.8-27B hits 381 tok/s on a single RTX 3090 via inference software alone — Redo · 2026-11-06
- Dolphin Networks details 3.8× throughput gains running DeepSeek-V4.1 Flash on PCIe GPUs — AlpinDale · 2026-11-06
- Lawyer runs DeepSeek-V4.1-Flash on 8x 2018 V100s: 215 tok/s on code, 524K context fully in VRAM — TumbleweedNew6515 · 2026-11-06
- Hyperkernel generates SOTA megakernels across 11 domains, now on GitHub — MillionInt · 2026-11-06
- Musk: space will be the cheapest place to run superintelligence in 2-3 years — XFreeze · 2026-11-06
- Ex-Microsoft engineer: Azure runs 1M+ VM services, adds a datacenter every 4 days — viksit · 2026-11-06