Skip to content

Nemotron 3.5 models arrive on Microsoft Foundry

Short answer

NVIDIA and Microsoft announced at Ignite that Nemotron 3.5 open models are available on Foundry Managed Compute, which now accepts custom model weights. A serverless post-training offering lets teams tune models on their own data. EY reports 55% higher accuracy and 15x lower cost on one classification task.

What this means for operators

If your team spends money on a proprietary model API for narrow, repetitive work such as classifying invoices, tagging tickets or extracting fields from contracts, this gives you a way to test a tuned open model for the same job. Build a small evaluation set from your own records, run the candidate against your current model, and compare accuracy and cost per document. The EY result comes from one purchasing-records task and its own dataset, so treat it as a reason to run the test, not as a forecast for your workload. The source gives no pricing, so cost has to be measured on your own volume.

NVIDIA says it is widening its work with Microsoft across cloud, edge and desktop, according to a post on the NVIDIA Blog published around Microsoft Ignite. For teams that automate business processes, the useful part is the model and tooling section.

What is new in Foundry

Microsoft Foundry gets a wider set of NVIDIA open models and customization options. Three pieces matter most:

  • Serverless post-training. Microsoft's Programmed Training offering lets developers steer how a model is tuned while Foundry runs the underlying infrastructure. The customer's job is to supply business data, check the model's answers and push behavior in the right direction.
  • Custom weights. Foundry Managed Compute can now deploy a customer's own model weights, which links tuning to hosting inside Microsoft's managed service.
  • Nemotron 3.5 lineup. Super is aimed at long-running agents doing coding, research and business workflows. Lightning targets applications that need quick responses and tool use. Bolt, built for high-throughput, low-latency reasoning, and Transcribe, a speech recognition model, are listed as coming soon.

Supporting models cover other parts of an agent pipeline: Embed and Rerank for finding relevant information, Parse for pulling structured data out of documents, and Content Safety for flagging harmful content.

The EY example

EY used Programmed Training to adapt Nemotron models to classify corporate purchasing records. On its own evaluation dataset, accuracy rose by 55% and cost fell 15x compared with proprietary frontier models. The source presents this as a demonstration that open models can be customized for specialized business tasks. It is a single vendor-reported case with no further detail on dataset size or method.

Running closer to your data

For organizations that want models near their own systems, Nemotron 3.5 support in Foundry Local brings reasoning, coding and tool-using agents to Azure Local environments. Azure Local will also support NVIDIA RTX PRO 6000 Blackwell GPUs from hardware partners, and it is extending to Jetson Thor and Jetson Orin edge platforms for industrial uses such as machine vision and robotics.

Infrastructure and desktop items

On the infrastructure side, Azure virtual machines built on NVIDIA Vera Rubin NVL72 systems are expanding, and the companies are working on confidential computing for bare-metal and rack-scale AI infrastructure so sensitive models and data can be protected during processing. For virtual desktops, Direct Stream capabilities are coming to GPU-accelerated Windows 365 Cloud PCs, supporting up to two 4K streams at 60 frames per second. A limited preview is open for sign-ups, with a public preview planned for early 2027.

What the source does not say

The post gives no prices, regional availability or plan requirements for the Foundry features. Anyone planning to adopt them should check Microsoft's own documentation for those details.

Source: NVIDIA Blog

INITE data

What we measured ourselves on the tools in this item. You will not find it in the source.

Data as of Nov 20, 2026

The tools in this story

Next step

Visibility Analyzer

An AI news item will not tell you how assistants see your own site. The Visibility Analyzer checks that on your live site: which answers cite you, which pages an engine cannot retrieve, and what to fix first. Free to run.

Run a free visibility audit

Free to run. No card.

Fee
Free
Length
One run, minutes

Free tier: two analyses a day, no card required.