NVIDIA Adds Super VL and Bolt Models to Nemotron 3.5 Family

Summary

NVIDIA has released Nemotron 3.5 Super VL, a 120-billion-parameter open-weight model designed for high-volume agentic workflows. The company also introduced two edge-focused Bolt models aimed at local and embedded deployment: the 74B-A9B, which uses 24 GB of memory, and the E4B, which uses 4 GB.

Why it matters

This model offers enterprise customers an open-weight, post-training tunable option that runs across the entire hardware stack, from DGX servers to RTX desktops. This reduces reliance on closed-source API endpoints for agent workloads. The emphasis on 2-bit quantization for local use, along with the release of open datasets and reinforcement learning environments, lowers the barrier for teams requiring on-premise and privacy-sensitive agent deployments.

Highlights

• Super VL claims a 5x higher data throughput speed compared to the second-best model in its class, according to the Artificial Analysis Intelligence Index.
• Bolt 74B-A9B fits into 24 GB of memory thanks to local 2-bit quantization; this enables the execution of local agentic tasks on RTX/DGX Spark.
• Open weights, training recipes, and reinforcement learning environments (NeMo Gym) have been published on Hugging Face for post-training fine-tuning.

Critical take

The '5x faster' claim in the headline was measured against a competitor group selected by the vendor in a single comparison. Furthermore, since the text is a vendor blog, it does not provide independent verification of data throughput or accuracy. The 2-bit quantization that allows Bolt 74B to fit into 24 GB is presented as a feature; however, the article does not quantitatively specify the quality loss compared to higher-precision references. This leaves a significant gap for practitioners evaluating real-world agent reliability.