On October 15, 2026, NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter model, alongside two edge models: Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB, extending its open model family from data center to desktop. The launch pushes high-performance always-on agent capabilities toward local devices, potentially reshaping open-source deployments in industrial inspection, robotics and personal assistants. Liu Gong believes NVIDIA is using an open model plus
On October 15, 2026, NVIDIA announced the expansion of its Nemotron 3.5 family through its official blog. The flagship Nemotron 3.5 Super VL is a 120-billion-parameter open model that NVIDIA says delivers the fastest task completion and highest accuracy for specialized always-on agents. Two edge models joined the lineup: Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB. The 74B-A9B variant has 74 billion total parameters, 9 billion active parameters and a 24GB memory footprint, optimized to balance intelligence and speed per GB for local agentic workloads. The E4B variant needs only 4GB, targeting lower-resource devices. NVIDIA did not disclose training data size, benchmark scores, license terms or specific hardware partners in the blog excerpt. The release extends the Nemotron open-source line from a 120B data-center tier to 24GB desktop and 4GB embedded tiers, creating a
NVIDIA's timing reflects a shift in open-source model competition from cloud-scale parameter counts to edge efficiency. Meta Llama, Mistral and Alibaba Qwen already cover the 7B to 70B range, but 120B-class open models remain scarce, and 70B-class models that run on 24GB of VRAM are rarer. Demand for local agents has risen sharply since 2025: industrial inspection requires low latency and data staying on-premises, personal assistants need offline always-on operation, and robots need real-time decision-making. NVIDIA has used the Nemotron series to demonstrate model capability, but its commercial goal is to keep developers on its GPUs, Jetson and RTX platforms, driving inference hardware
Liu Gong's take: within six months, at least one industrial inspection or robot patrol vendor will ship a local agent built on Nemotron 3.5 Bolt, because the 24GB and 4GB memory footprints map directly to existing desktop GPUs and embedded devices. From a technical standpoint, the 120B Super VL targets data-center always-on agents, while Bolt 74B-A9B uses a sparse structure with 74B total and 9B active parameters to compress inference into 24GB. As an AI visual inspection practitioner, Liu Gong notes that industrial quality-check sites often have only a single 24GB GPU. Deploying a 70B model there previously required 4-bit quantization and still produced unstable latency; if Bolt's activation sparsity truly optimizes intelligence per GB, it would be the first time a 70B-class model runs a stable always-on task on one card. On the business side, NVIDIA's real goal is to sell hardware through models: Super VL attracts cloud providers buying H-series GPUs, while Bolt drives RTX workstations and Jetson sales. Open models do not directly generate revenue, but they create CUDA path dependency. For the industry, Meta Llama and Alibaba Qwen now face a competitor with deep chip-level optimization. NVIDIA can co-tune model, driver and inference library in ways pure model companies cannot. Liu Gong predicts that within 12 months, mainstream edge AI boxes will add "supports Nemotron 3.5 Bolt" as a selling point, similar to earlier "supports Llama" labels. The risk is that NVIDIA has not published benchmark scores or license details; if the open license restricts commercial use or real speed falls short, community migration will suffer.
- Nemotron 3.5 Super VL has 120 billion parameters, about 1.6 times the 74B total parameters of Bolt 74B-A9B, positioning them for data-center and local always-on agents respectively.
- Bolt 74B-A9B needs 24GB of memory, matching a single RTX 4090/5090-class desktop GPU; a typical 70B model requires about 140GB in FP16, and even 4-bit quantization still needs roughly 35-40GB, so 24GB depends on a combination of activation sparsity and quantization.
- Bolt E4B needs only 4GB, creating roughly a 30x parameter-scale gap with the 120B Super VL and targeting embedded terminals with 8GB or less of memory.
- NVIDIA did not disclose a specific benchmark improvement percentage over the previous
For the industry, in the short term within three months, the open-source community will quickly fine-tune vertical models for industrial, medical and robotics use on Nemotron 3.5 Bolt, lowering the barrier to edge agent development. In the medium term within one year, NVIDIA may bind model performance to its hardware through CUDA and TensorRT co-optimization, squeezing pure model companies on the edge. For ordinary users, in the short term a 24GB local assistant can run on a high-end desktop PC, offering better privacy and response speed than pure cloud solutions. In the medium term, personal agents may become resident on laptops and phones, with 4GB models making offline voice and visual
Questions worth exploring:
- How NVIDIA uses open models to drive GPU and Jetson hardware sales, forming a software-hardware ecosystem loop.
- Whether Bolt 74B-A9B's activation sparsity can let a 70B-class model run stable always-on agents on a single 24GB desktop GPU.
- The competitive impact of a 120B open model on Meta Llama, Alibaba Qwen and open-source licensing strategies.
Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.
This item comes from NVIDIA Blog (published 2026-10-15 15:00:24). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.