#

Key takeaways

• NVIDIA expanded its Nemotron 3.5 open-model family on October 15, 2026.

• The update adds a 120-billion-parameter vision-language model and two Bolt models for local workloads.

• NVIDIA says some models are available now, while one 4GB model is scheduled for October 23.

• The company links the release to different speed, reasoning, and deployment profiles.

• NVIDIA also points to open weights and customization through NVIDIA NeMo.

NVIDIA expands Nemotron 3.5

NVIDIA announced an expansion of its Nemotron 3.5 open-model family on October 15, 2026. The update adds new models for different agentic workloads and hardware limits. It also extends the lineup beyond the existing Lightning and Ultra variants.

The main addition is Nemotron 3.5 Super VL. NVIDIA describes it as a 120-billion-parameter vision-language model. It is aimed at specialized agents that process both visual and text information. That makes it the most clearly multimodal model in the new set.

NVIDIA also introduced two Bolt models. Nemotron 3.5 Bolt 74B-A9B 24GB is designed for local agentic workloads. Nemotron 3.5 Bolt E4B 4GB is aimed at small local systems and conversational or physical agents. The naming suggests different hardware targets, but the source only supports the stated deployment focus.

Availability and deployment profile

NVIDIA says Super VL and the 24GB Bolt variant are available now. The 4GB Bolt variant is scheduled for October 23. It should not be described as already generally available. The staged rollout matters because it separates immediate access from a later release date.

The company positions these additions alongside Lightning and Ultra. That gives developers more choice across speed, reasoning, and deployment profiles. In practical terms, the family now spans larger multimodal systems and smaller local models. The source frames this as a design choice for different agent needs, not as a claim that one model fits every task.

Performance claims and benchmarks

NVIDIA makes performance claims in the announcement. It says the models deliver faster completion and high accuracy in the comparisons it chose. Those results should be attributed to NVIDIA and to the relevant benchmarks. They should not be generalized to all production workloads.

That distinction matters. Benchmark results can show how a model behaves under specific test conditions. They do not automatically prove the same outcome in every deployment. The source supports the benchmark framing, so the claims should stay tied to NVIDIA's own comparisons.

Open weights and customization

NVIDIA says the models come with open weights. It also points to customization resources through NVIDIA NeMo. That suggests the company wants developers to adapt the models rather than use them only in a fixed form.

The source does not provide technical implementation details beyond that. It does not describe training data, fine-tuning steps, or integration requirements. So the safest reading is that NVIDIA is emphasizing flexibility, not publishing a full deployment guide.

Reported areas of interest

The announcement cites teams in finance, cybersecurity, scientific discovery, and healthcare administration as exploring Super VL. These examples describe intended or reported use. They are not verified operational deployments in Morocco or elsewhere in the supplied material.

The list does show the kind of work NVIDIA expects these models to support. It also reinforces the multimodal angle, since such domains can involve both text and visual inputs. Still, the source only supports interest or exploration, not confirmed outcomes.

Morocco relevance

The source reports no Morocco-specific fact. For readers, the global lesson is simple: model families can now split between large multimodal systems and smaller local options. That can help teams match model size to hardware limits and task needs, while keeping benchmark claims tied to the vendor's own tests.

Bottom line

This release is less about one model and more about a broader family strategy. NVIDIA is expanding Nemotron 3.5 across larger vision-language use cases and smaller local agentic systems. It is also keeping the availability timeline explicit, which helps readers separate what is live now from what is still scheduled.