In-depth analysis · About 5 min read
Nemotron 3.5 separates multimodal agents from small local deployments
Super VL, a 120B-parameter model, arrives alongside a 24GB Bolt variant; performance and availability claims depend on model and configuration.
- Source published
- Published here
- Source checked
NVIDIA released Nemotron 3.5 Super VL and Bolt 74B-A9B 24GB on October 15. A smaller Bolt E4B 4GB model is scheduled for October 23 and, as of October 20, remains a future release. The announcement covers one model-family expansion, with distinct server and local uses. It does not mean that the 120B Super VL model fits into Bolt's stated 24GB footprint.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- Super VL has 120 billion total parameters and activates 12 billion per token, according to NVIDIA's technical description. Active parameters are not the full model's memory requirement.
- NVIDIA reports visual-reasoning and inference results under specific conditions. Its Dynamo example uses 16 B200 GPUs at 50 tokens per second per user; it is not a result for a single consumer GPU.
Three variants have different sizes and release states
Super VL is a mixture-of-experts model with 120B total parameters and 12B activated per token. That design reduces the computation used for a token, but it does not turn the whole checkpoint into a 12B-parameter download. The concurrently released Bolt 74B-A9B 24GB targets local agent work. Bolt E4B 4GB targets smaller conversational and physical-agent applications, with availability announced for October 23. These models join the broader Lightning and Ultra family. Their different roles and dates should remain visible when comparing capabilities or planning a deployment.
What the visual capabilities are intended to handle
The focus is the material agents encounter in work: OCR, tables, charts, webpages, PDFs, images and video. A unified model can retain context instead of routing text and visual tasks between separate systems. NVIDIA reports 80.8% on CharXiv, rising to 84% when Python code interpretation is available. The second number includes a tool condition and is not a higher score for the unchanged tool-free setup. Our reading is that the useful evaluation unit is the whole workflow: document quality, tool access and answer verification all affect whether extracted information can support a decision.
Customisation includes training methods, not just a prompt
The release includes weights, datasets and training recipes, along with reinforcement-learning environments. The technical account describes supervised fine-tuning and LoRA through NeMo AutoModel and reinforcement learning through NeMo RL and Gym. NVIDIA names applications including credit-document analysis and coding agents as examples of partner customisation. Such examples show intended specialisation, not independent proof that every customer receives the same improvement. For a team considering this route, domain data, evaluation cases and permissions to use that data are separate prerequisites; open availability does not remove the need to read each applicable licence.
Why an efficiency figure needs its serving context
NVIDIA attributes throughput gains to several layers. NVFP4 uses four-bit weights and activations on Blackwell, while non-Blackwell execution uses four-bit weights with 16-bit activations. Quantisation-aware distillation aims to preserve behaviour at reduced precision. Multi-token prediction drafts several tokens for verification, and prefix caching reuses shared instructions, tools and history. The reported throughput of up to five times comparable models in its class is therefore a serving result, not a guaranteed reduction in every task's completion time. We have not tested these techniques or measured their interaction on another configuration.
The 24GB local claim belongs to Bolt
Bolt 74B-A9B uses native two-bit quantisation and is described as fitting in 24GB of memory for local agent workloads on RTX and DGX Spark systems. That is a particular model configuration, not a memory statement about Super VL or a complete system purchasing specification. The smaller E4B carries a 4GB designation but still has its later release date. In our assessment, local deployment should be checked with the intended context length, other running applications and tool behaviour. Local model inference also does not automatically make an agent's web searches or connected tools offline.
Deployment choices and a tightly specified Dynamo example
Super VL is offered through Hugging Face, ModelScope, OpenRouter and NVIDIA NIM; Bolt has routes including llama.cpp, Ollama, LM Studio, Unsloth and vLLM. NIM supplies validated configurations, with enterprise support available through NVIDIA AI Enterprise. The technical blog's Dynamo example uses 16 B200 GPUs at 50 tokens per second per user: cache-aware routing and offload serve up to eight times as many users while reducing first-token latency by 85%. Those hardware and service conditions matter. The result is neither an eightfold improvement for every laptop nor evidence of a particular subscription's included usage or price.
Why it matters
Teams gain a self-hostable and customisable option for document-centred agents, plus a separate small-memory local path. The practical choice requires a representative workflow, its tool permissions and its actual serving hardware, rather than a single headline speed figure.
Limits and availability
We have not run these models or reproduced the benchmarks. The sources do not establish universal commercial licence rights, an all-in hardware requirement, a hosted API price or identical quality across every deployment route.
Sources and evidence
Source published: Oct 15, 2026 · Source checked: Oct 20, 2026
Further reading
Source checked: Oct 20, 2026
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.