AitopkitNVIDIA has released Nemotron 3.5 Super VL on October 15, 2026 — a 120-billion-parameter open multimodal mixture-of-experts model (12B active per token) built for always-on enterprise agents. It reads text, images, documents, and short videos, then reasons, calls tools, and writes code in a single model. Weights ship under the OpenMDW-1.1 license, which permits commercial use.
The headline result: Super VL scores 26.5 on the Artificial Analysis Intelligence Index, making it the leading US open-weight model — ahead of Thinking Machines' Inkling Small (25.7) and NVIDIA's own Nemotron 3 Ultra (23). NVIDIA also claims up to 5x higher single-GPU throughput than open peers, and the NVFP4 checkpoint scores nearly identically to BF16. Context stretches to 1 million tokens.
Two edge siblings landed the same day: Nemotron 3.5 Bolt 74B-A9B, a 74B-total/9B-active model quantized to 2 bits so it fits in 24GB of memory — a single RTX 5090-class machine — scoring 21 on the Intelligence Index; and a 4GB Bolt E4B for device agents, due October 23.
NVIDIA also released open datasets, training recipes, and RL environments in NeMo Gym alongside the weights — a serious play for the agentic-AI developer stack, not just a model drop.
The catch is the usual one: benchmark scores measure the evaluator's workload, not yours. Test document accuracy, tool use, and deployment controls against your own pipeline before standardizing on it.