NVIDIA open-sources Nemotron 3.5 Super VL: 120B MoE with 5x faster agentic throughput
NVIDIAAI · x · 2026-10-15
NVIDIA released and open-sourced Nemotron 3.5 Super VL, a multimodal model detailed in a new technical blog.
- Architecture: 120B-parameter MoE with 12B active per token, unifying reasoning, instruction following, tool use, and text/image/document/video understanding in one model.
- Performance: up to 5x higher agentic task throughput on a single GPU vs. same-size open models; retains strong reasoning at 100K input tokens.
- Inference optimizations: LiLiCorr speculative decoding, NVFP4 quantization with quantization-aware distillation, multi-token prediction, and context reuse.
- Adoption: Scale AI, Ramp, and AI21 post-trained it for specialized workflows, seeing up to 40x lower inference cost, 10x lower latency, and 40% higher accuracy.
- Openness: weights, data, and recipes are public on Hugging Face; customizable via NeMo (SFT, LoRA, RL).
More from Infra
- Muse-Spark open-weights drop: 816B-A42B MoE model gets 359GB MLX 3-bit build — ns123abc · 2026-10-16
- vLLM ships Day-0 LoRA RL for Muse Spark Open, lifting AIME from 0.56 to 0.85 — vllm_project · 2026-10-16
- DSpark draft model for Muse Spark Open proposes 7 tokens per pass with 23-42% acceptance — vllm_project · 2026-10-16
- Cloudflare Stream's WebRTC ultra-low-latency live streaming is now Generally Available — ritakozlov · 2026-10-16
- YC-backed Overshoot launches inference optimization, claims $5M saved for AI teams — ycombinator · 2026-10-16
- Meta's 816B open-weight Muse Spark Open hits Groq preview at 350+ tokens/sec — GroqLLC · 2026-10-16