NVIDIA unveils Nemotron 3.5 Super and Lightning: open MoE hybrid Mamba-Transformer models
kalyan_kpl · x · 2026-10-30
A paper introduces NVIDIA's Nemotron 3.5 Super and Nemotron 3.5 Lightning—open, efficient Mixture-of-Experts hybrid Mamba-Transformer models built for reasoning, long-context understanding, and agentic workloads.
- Super: 120B total / 12B active multimodal; Lightning: 30B total / 3B active text-only.
- Both combine large-scale distillation, long-context training, and native speculative decoding for efficient deployment.
- Super adds Indic continued pretraining and agentic software-engineering mid-training.
- Post-training uses SFT plus RL (RLVR and RLHF), with specialized teacher training and multi-teacher on-policy distillation (MOPD) in Super.
Related event: NVIDIA Open-Sources Nemotron 3.5 MoE Mamba-Transformer Models(2 posts)→
More from Models
- Mistral Large 4 matches GLM on coding with 47% fewer tokens, halves preview price — petrusenko_max · 2026-10-31
- Mistral Large 4 matches GLM-5.3 at 48% on coding agent index with 47% fewer tokens — petrusenko_max · 2026-10-31
- Grok on LLM identity: why no model ever identifies as a web server — mimi10v3 · 2026-10-31
- Intelligence per dollar fell ~1,000,000x in three years, shows double-exponential chart — banteg · 2026-10-31
- Altman: each math AI proof is worth $10M, and cost will drop 100x within a year — Traditional-Chip8339 · 2026-10-31
- GPT high-tier model accused of task-abandoning and sycophancy in coding and lyrics — Kyrannio · 2026-10-31