NVIDIA's Nemotron 3.5 Super hits 51.78% on SWE-bench Pro with 4.6x fewer active params
dair_ai · x · 2026-10-30
NVIDIA released a technical report for Nemotron 3.5 Super and Lightning, open MoE hybrid Mamba-Transformer models for agentic AI:
- Super: 120B total / 12B active, multimodal (text/image/video), scores 51.78% on SWE-bench Pro, beating Nemotron 3 Ultra with 4.6x fewer active parameters
- Lightning: 30B total / 3B active text-only model
- Pipeline includes agentic software-engineering mid-training, SFT, RL (RLVR/RLHF) and multi-teacher on-policy distillation; Indic continued pretraining lifts Indic benchmarks by up to 9.82 points
- Serving: up to 4,856 output tokens/s/GPU on GB200 while sustaining ≥50 tokens/s/user; NVFP4 quantization recovers 99.62% median accuracy vs BF16 with 2.43x throughput, and delivers 2.34x/1.90x the throughput of Qwen3.5/Qwen3.8 Flash-Next
- Checkpoints (base, BF16, NVFP4), datasets, recipes, and speculative drafters are open-sourced
Related event: NVIDIA Open-Sources Nemotron 3.5 MoE Mamba-Transformer Models(2 posts)→
More from Infra
- Your AI agent would cost 260% more without prompt caching, Arena data shows — arena · 2026-10-30
- LiteLLM flaw let attackers pull upstream API keys from Postgres, sparking gateway key storage debate — Early-Sir-932 · 2026-10-30
- Nebius runs auction-driven GPU pricing, lifting rates ~20% amid tight supply to 2028 — BenBajarin · 2026-10-30
- macOS update silently re-adds eMudhra root certs users had removed — steipete · 2026-10-30
- Bittensor's Terza Inference routes one API across miners' models on subnet 103 — const_reborn · 2026-10-30
- SGLang Summit 2026 sells out sponsor booths and Full Passes ahead of Nov 12–13 event — ying11231 · 2026-10-30