NVIDIA announced three new open‑source Nemotron 3.5 models: the 120‑billion‑parameter Super VL, the 74‑billion‑parameter Bolt 74B‑A9B (24 GB), and the 4‑billion‑parameter Bolt E4B (4 GB). All three are positioned as drop‑in replacements for existing agentic workloads, ranging from enterprise servers to edge devices, and they are released with open weights, datasets, and training recipes.
Model Portfolio and Performance Claims
Super VL is described as a “frontier‑accuracy” model that matches the capability of NVIDIA’s largest reasoning model while using roughly one‑fifth the parameters. The vendor claims it is up to five times faster than the next best open model in its class and leads the Artificial Analysis Intelligence Index leaderboard for throughput. Bolt 74B‑A9B uses native 2‑bit quantization to fit within a 24 GB memory envelope on RTX/DGX Spark hardware, offering a balance of intelligence and speed that the source calls the #1 open model on PinchBench and #2 on BrowseComp. Bolt E4B targets 4 GB devices such as GeForce RTX, Jetson, and RTX/DGX Spark, focusing on conversational agents and fast local voice assistants.
Deployment and Operational Considerations
All three models are open, meaning organizations can host them on any NVIDIA‑compatible infrastructure without sending data to a cloud provider. Super VL’s size (120 B) and Bolt 74B‑A9B’s 24 GB memory requirement imply that deployment will need GPUs with at least that amount of VRAM, while Bolt E4B can run on much smaller edge devices. The 2‑bit quantization of Bolt 74B‑A9B reduces memory pressure but may require compatible runtime libraries that understand the format. Because the models are open, teams can integrate them into existing CI/CD pipelines, store the weights in artifact repositories, and apply the provided training recipes for post‑training on proprietary data.
Security and Governance Implications
Open weights and datasets give engineers full visibility into the model supply chain, supporting audit and compliance requirements. Running the models on‑prem satisfies data‑residency constraints and eliminates the need to transmit potentially sensitive prompts or outputs to external services. However, the ability to post‑train on internal data introduces a responsibility to secure the fine‑tuning pipeline, manage access to the training data, and validate that any downstream model artifacts do not leak proprietary information. Organizations should treat the model artifacts as sensitive assets, applying the same secret‑management and access‑control practices used for other binaries.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
- Validate hardware compatibility: ensure GPUs meet the VRAM requirements for Super VL (≥24 GB) and Bolt 74B‑A9B, and confirm that edge devices can run 2‑bit quantized models.
- Benchmark throughput against existing agents to confirm the claimed 5× speed advantage for Super VL and the PinchBench ranking for Bolt 74B‑A9B.
- Integrate model weights and training recipes into version‑controlled artifact stores; treat fine‑tuned models as production binaries.
- Design fine‑tuning pipelines with data‑privacy controls, logging, and audit trails to meet governance policies.
- Plan for monitoring resource usage and latency on edge deployments, especially for Bolt E4B where memory headroom is limited.
By mapping the new Nemotron variants to the appropriate hardware tier and embedding the open‑model workflow into existing DevOps processes, teams can gain higher‑throughput agentic capabilities while retaining full control over data and security posture.
