NVIDIA has placed another heavy bet on the open-source large model赛道. In mid-October, NVIDIA officially launched Nemotron 3.5 the Nemotron 3.5 series, unveiling three open-weight models at once—Super VL（120B）Nemotron 3.5 Super VL, targeting enterprise-grade Agents,Bolt 74B-A9BNemotron 3.5 Edge, aimed at local AI PCs, and the smallest, Bolt E4BNemotron 3.5 Chat, dedicated to lightweight dialogue. This marks NVIDIA’s most comprehensive product lineup yet in the open-source Agent space, following Nemotron 3 Ultra Nemotron 3 Ultra. Its core value proposition boils down to one sentence:Enable organizations to run their own agents at any compute tier—no trade-off between “capability” and “deployability.”.
Core capabilities: Three models spanning enterprise to individual use
The biggest shift in the Nemotron 3.5 series is moving away from the previous “one flagship dominates all” approach to three clearly differentiated lines.
Nemotron 3.5 Super VL（120B）Nemotron 3.5 Super VL is this round’s flagship, designed for enterprise scenarios requiring “always-on” Agents. Though its parameter count is significantly smaller than the 550B Nemotron 3 Ultra, NVIDIA claims it delivers comparable accuracy ininstruction following and tool calling,two critical capabilities. Official target use cases include document intelligence, financial accounting, scientific experimentation, health insurance claims processing, and security offensive/defensive resilience—precisely the tasks enterprises are most willing to pay for under strict “data stays within domain” requirements. NVIDIA also cites a roster of heavyweight customer endorsements: Skydance, EY, Ramp, CrowdStrike,Perplexity and Salesforce—all deploying Super VL.
Nemotron 3.5 Bolt 74B-A9Brepresents the “local-first” approach. It is a 74B-parameter Mixture of Experts (MoE) model with approximately 9B activated parameters, specifically designed for local agentic tasks, aiming tostrike a balance between speed and reasoningand handle multi-step reasoning, tool calling, and web browsing—workflows requiring sequential decision-making. It runs on next-generation AI PCs equipped with NVIDIA RTX Spark chipsand DGX Spark workstations—devices now being rolled out at scale by partners including Microsoft and Dell, targeting consumers, programmers, and enterprise knowledge workers.
The smallest Bolt E4Bis positioned as a conversational agent, with fewer parameters and lower power consumption, suitable for lightweight voice or text interactions. Super VL and Bolt 74B-A9B are already available for download; Bolt E4B will be released later this month.
User experience and limitations
As an open-weight model, Nemotron 3.5’s biggest appeal lies in“data sovereignty”NVIDIA Software Vice President Kari Briski put it plainly: unlike closed frontier models, organizations can perform post-training of Super VL on their own data, deploy it on their own infrastructure, and truly “own what they build.” For compliance-sensitive domains such as healthcare, finance, and government, this capability is irreplaceable by closed-source APIs.
On performance, NVIDIA has released notably aggressive benchmarks: after task-specific post-training for agentic workloads, Super VL’s task completion speed surpasses Anthropic’s Opus 5 and OpenAI’s GPT-6.1 Sol, Luna, and Astra. However, such “official self-benchmarks” are already industry-standard practice, and real-world effectiveness awaits third-party evaluations and practical deployment validation.
Limitations are equally apparent:
- Deployment barrier: Open source ≠ plug-and-play. Super VL has 120B parameters and requires enterprise-grade multi-GPU setups for local inference; while Bolt 74B runs on AI PCs, its 24GB VRAM requirement still excludes many mainstream users.
- Chinese-language capability remains questionable: Official materials do not emphasize Chinese-language performance; domestic users targeting Chinese writing or Chinese-language agents will likely find domestic open-source models more convenient.
- Ecosystem is still maturing: Compared to CUDA’s dominance in training, Nemotron’s inference ecosystem, community tooling, and fine-tuning tutorials require further time to mature.
Nemotron 3.5 Super VL vs. peer open-source models: horizontal comparison
| Model | Parameter count | Core Positioning | This Site’s Rating |
|---|---|---|---|
| Nemotron 3.5 Super VL | 120B | Enterprise Agent On-Premises | 8.3 |
| Nemotron 3 Ultra | 550B | Strongest open-source model in the U.S. | 6.8 |
| GLM-5.3 | Open-Source Inference | Programming Agent | 8.4 |
| DeepSeek V4 Pro | MoE Flagship | Flagship at Mass-Market Pricing | 8.9 |
| Qwen3.8 2.4T | 2.4T | Tongyi Open-Source Milestone | 8.5 |
| MiniMax M3 | Open Weights | Million-Token Programming | 8.1 |
From a coordinate-system perspective, Nemotron 3.5 Super VL fills NVIDIA’s gap in the “mid-to-large enterprise Agent” niche: it is more pragmatic than its predecessor Nemotron 3 Ultra (half the parameters, easier deployment), yet DeepSeek V4 Pro,Qwen3.8 Compared with these domestic open-source flagships, Chinese-language ecosystem support and community momentum remain weaknesses. If you are an overseas enterprise prioritizing data sovereignty and hardware integration, Super VL offers unique value; if you are a domestic team building Chinese-language Agents, domestic models remain the safer choice for now.
Overall Score
| 维度 | Score | evaluate |
|---|---|---|
| functional completeness | 8.5 / 10 | Comprehensive tool calling, multi-step reasoning, and enterprise-task coverage |
| 易用性 | 7.5 / 10 | High open-source deployment barrier; local execution requires multi-GPU setup |
| Cost-effectiveness | 8.5 / 10 | Weights are free, but hardware and operational costs are self-managed |
| 中文支持 | 7.0 / 10 | Official Chinese-language optimization is limited; domestic use-case performance is suboptimal |
| 输出质量 | 8.5 / 10 | Post-training on agentic tasks achieves performance comparable to cutting-edge closed-source models |
Overall rating: 8.3/10
Frequently Asked Questions (FAQ)
Is Nemotron 3.5 Super VL free?
Yes—the model weights are open-source and freely downloadable, and may be freely fine-tuned and used commercially. However, “free” applies only to the model itself; actual operation requires self-provided GPU hardware, plus electricity and operational expenses—enterprise-scale deployment of this 120B-parameter model incurs significant cost.
What distinguishes Nemotron 3.5 Super VL from Nemotron 3 Ultra?
Super VL has fewer parameters (120B vs. 550B), yet NVIDIA claims its accuracy in instruction following and tool calling is comparable; it is positioned as a more pragmatic, easier-to-deploy enterprise Agent model, whereas Ultra pursues maximum scale as a flagship. For most enterprises, Super VL delivers better cost-performance balance.
What configuration is required to deploy Nemotron 3.5 locally?
Bolt 74B-A9B targets devices with 24 GB VRAM and runs on AI PCs equipped with RTX Spark chips or DGX Spark workstations; the flagship Super VL model has 120 billion parameters and requires enterprise-grade multi-GPU clusters.
Does Nemotron 3.5 support Chinese? How well does it perform?
The official documentation does not emphasize Chinese-language capabilities, and its optimization for Chinese-language scenarios is likely inferior to domestic open-source models such as GLM andDeepSeekQwen. For Chinese-language writing or Chinese-language Agents, domestic users are advised to prioritize leading domestic open-source models.
Nemotron 3.5 and DeepSeekQwen—which should you choose?
It depends on the use case. For enterprise Agent privatization with emphasis on data sovereignty and NVIDIA hardware integration, choose Super VL; for Chinese-language programming and Chinese-language Agents,DeepSeek V4 Pro and Qwen3.8 Qwen remains the more mature choice.
Summarize
The significance of the Nemotron 3.5 series extends beyond adding another open-source model—it marks NVIDIA’s transformation of “open-source Agents” from a slogan into a complete product matrix spanning enterprises, workstations, and personal PCs. For teams seeking private Agent solutions without vendor lock-in to a single cloud provider, this is a compelling new option to include in their evaluation list.
Want to discover more useful AI models and tools? Check out our AI Model Library and Tool Comparison Engineor continue reading:Nemotron 3 Ultra Deep Review · GLM-5.3 Deep Review · Mistral Large 4 Release.