Technology
October 28, 2026 1078 words

NVIDIA Nemotron 3.5: Open Models for Agents, Speech and Enterprise Search

Hams AI Tech
Professional Content Team
Expert in AI Prompts & Professional Tools

NVIDIA is best known for the chips that train and run AI, but it has also become a major publisher of open models. During October 2026, it expanded the Nemotron 3.5 family with models for always-on agents, local workloads on personal hardware, speech transcription and enterprise search. Taken together, these releases give organizations a set of open building blocks that run well on NVIDIA hardware, from laptops with RTX graphics to data center servers.

Nemotron 3.5 Super VL

Nemotron 3.5 Super VL is an open multimodal model with 120 billion parameters, of which 12 billion are active per token. NVIDIA designed it for always-on agents, systems that run continuously, watching inboxes, documents and data feeds and acting when needed. The low active parameter count keeps running costs manageable for that kind of continuous workload.

Super VL has already been adopted by partners. ElevenLabs announced at its November summit that Nemotron 3.5 Super VL is available inside ElevenAgents, its platform for voice agents.

Nemotron 3.5 Bolt: Agents on Your Own Hardware

Bolt is the family's local option. Nemotron 3.5 Bolt 74B-A9B in a 24 gigabyte configuration targets local agent workloads on RTX graphics cards and DGX Spark. A smaller Bolt E4B in a 4 gigabyte footprint followed later in October for lighter devices.

Running models locally matters for professionals who handle confidential information. A model that runs on a workstation in your office never sends client documents to an external server. Our guide to running AI locally explains the trade-offs, and our article on the DGX Spark Developer Kit covers NVIDIA's desktop hardware.

Nemotron 3.5 Transcribe

Announced on October 28, Nemotron 3.5 Transcribe is an open, promptable speech model with about 34 billion total parameters and 4.3 billion active. It uses a hybrid Mamba-Transformer mixture-of-experts design and handles several tasks in one checkpoint:

    • Transcription in 40 languages.
    • Speaker attribution.
    • Translation.
    • Audio question answering, on recordings up to an hour long.

You choose the transcript style with a plain-language instruction, for example verbatim, clean or summarized. Weights are available on Hugging Face and ModelScope in BF16, FP8, NVFP4 and GGUF formats under the OpenMDW-1.1 license, with day-one support in vLLM and SGLang and hosted endpoints on Baseten and DeepInfra. NVIDIA also released the open Granary v2 data pipeline for fine-tuning.

For law firms, medical practices and consultancies that record meetings, a promptable transcription model that can run privately is a significant option. See our voice AI roundup for comparisons.

Nemotron 3.5 Embed, Rerank and Parse

At GTC Berlin, NVIDIA released three models aimed at grounding enterprise agents in company data:

    • Embed: creates searchable representations of text and page images.
    • Rerank: orders search results by relevance.
    • Parse: extracts structure from documents.

These models power retrieval systems, the technology that lets an AI assistant answer questions using your own files rather than general knowledge. NVFP4 variants are available for efficient inference. If your organization is building an internal knowledge assistant, these three models cover the core retrieval pipeline.

Why NVIDIA Publishes Open Models

NVIDIA's strategy is straightforward: open models that run well on NVIDIA hardware increase demand for that hardware. For customers, the result is a growing catalog of free, high-quality models optimized for the chips they already own. The same logic explains NVIDIA's investments in inference software such as llm-d, covered in our article on AI developer infrastructure.

Practical Uses for Professionals

Private Knowledge Assistants

Combine Embed, Rerank and Parse with a local or private model to build an assistant that answers questions about your policies, precedents or procedures. Law firms can complement this with structured tools like the Legal Research Notion template.

Meeting and Interview Transcription

Nemotron Transcribe can turn client meetings, depositions or interviews into searchable transcripts. Always obtain consent before recording. Explore legal document resources for downstream workflows.

Always-On Monitoring Agents

Super VL can watch data feeds and flag issues. Finance teams might use it to monitor transactions, while operations teams track system alerts. See the tax liability forecasting engine for a finance example and AI automation resources for more.

Getting Started

    • Identify whether your use case needs local deployment for confidentiality.
    • For local work, test Bolt on an RTX workstation or DGX Spark.
    • For transcription, run Nemotron Transcribe on a sample of real recordings and compare accuracy.
    • For knowledge assistants, start with a small document set and measure answer quality before scaling.
    • Review the OpenMDW-1.1 and other license terms before commercial deployment.

Building a Private Stack Step by Step

A common question from firms is how all these pieces fit together. A practical private stack might look like this. Parse converts your documents, including scanned pages, into structured text. Embed turns that text into searchable vectors stored in a database on your own servers. When someone asks a question, the system retrieves candidate passages, Rerank orders them by relevance, and a generator model such as Bolt or Super VL writes an answer that cites the passages it used. Transcribe can feed meeting recordings into the same pipeline, so decisions made in calls become searchable alongside written records.

Each component can be swapped later without rebuilding the whole system. That modularity is the real advantage of open models: you can start small on a single workstation, prove value on one practice area, then scale to a server as usage grows, all while keeping data under your control.

Frequently Asked Questions

Are Nemotron models free?

The weights are openly available. Costs come from the hardware or hosted services you use to run them.

Can Nemotron run without NVIDIA hardware?

Some formats, such as GGUF, run on other hardware through community tools, but performance is optimized for NVIDIA GPUs.

Which Nemotron model should a small firm start with?

For local experiments, Bolt is the most accessible. For transcription, start with Nemotron Transcribe through a hosted endpoint before considering self-hosting.

Does Nemotron Transcribe support translation?

Yes. It handles transcription, speaker attribution, translation and audio question answering in one model.

The Bottom Line

The Nemotron 3.5 family gives organizations open, practical tools for agents, speech and search, optimized for hardware many already own. For privacy-conscious professionals, it is one of the strongest foundations for building AI that stays in-house. Browse our agentic AI resources and technology AI tools to plan your deployment.

Related AI News

Topics & Keywords
NVIDIA Nemotron 3.5 Nemotron Super VL Nemotron Bolt Nemotron Transcribe open models RTX DGX Spark