NVIDIA released the open-source model Nemotron 3.5 Lightning (30B) and NeMo Switchyard for agentic AI
On 11 August, NVIDIA released the open-source model Nemotron 3.5 Lightning (30 billion parameters) for agentic AI tasks such as code review or tool use, under the OpenMDW 1.1 license with training data, alongside NeMo Switchyard for routing between models.
On 11 August, NVIDIA released Nemotron 3.5 Lightning, a mixture-of-experts model with 30 billion parameters designed for specialized tasks within larger multi-agent systems — according to the company, it is the most efficient model in its class for long-running agentic tasks (long-running agentic AI workloads). The model is intended for tasks such as code review, tool use, monitoring security alerts or answering billing questions. It is smaller than the existing Nemotron 3 Ultra (550 billion parameters) and larger than Nemotron 3 Nano (8 billion), and expands the Nemotron 3 family, which NVIDIA introduced last December. Weights, code, training recipes and training data were released under the Linux Foundation OpenMDW 1.1 license, which, according to sources, makes it a truly open model; deployment targets include local devices such as RTX PC, DGX Spark and Jetson.
Alongside the model, NVIDIA also released the NeMo Switchyard library, an open tool for routing requests within agentic applications. According to the company, it allows a specific request to be routed to the most suitable model across a custom mix of open models, proprietary models and models from NVIDIA.
The release came less than a month after NVIDIA CEO Jensen Huang voiced support for open models on X while the US government was considering restrictions on open Chinese models; according to Huang, open models strengthen sovereignty and improve security. Nemotron 3.5 Lightning is the first open model from NVIDIA since discussion of Chinese open models intensified following the release of low-cost, powerful models from Alibaba and Moonshot AI; it came out the day after the latest open model from Meta. Analyst Bradley Shimmin (Futurum Group) said that many consider Qwen 3.8 Max from Alibaba a benchmark for local agentic development, and welcomed the fact that NVIDIA also released the training data. Analyst Arun Chandrasekaran (Gartner) added that open models help NVIDIA sell its own infrastructure because running models requires networking, inference software and data training.
Why it matters
The model allows agentic tasks to run locally (on RTX PC, DGX Spark or Jetson) instead of through a paid API, addressing two common problems — high and hard-to-predict token costs and requirements for data control or sovereignty during deployment. The full release of weights, code, training recipes and data under the OpenMDW 1.1 license also allows companies to verify what data the model was trained on, which is relevant to assessing the risks of AI deployment. The NeMo Switchyard library addresses a practical problem in agentic systems that combine multiple models — automatically routing a request to the most suitable model from the available mix.
Release card
Nemotron 3.5 Lightning
NVIDIA
- Specifications
- 30 billion (mixture-of-experts)
- Licence
- Linux Foundation OpenMDW 1.1
- Availability
- Weights, code, training recipes and training data are freely available to download under the Linux Foundation OpenMDW 1.1 license; the model can run locally on NVIDIA RTX PC, DGX Spark and Jetson devices.
- code review within multi-agent systems
- tool use and routing requests between models (with the NeMo Switchyard library)
- monitoring security alerts and answering billing questions
According to NVIDIA, it is the most efficient model in its class for long-running agentic AI tasks within multi-agent systems.
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
What was added since the original report
Verified updates
-
The model has 30 billion parameters (specific size); Released under the Linux Foundation OpenMDW 1.1 license with weights and training recipes; Accompanying tool NeMo Switchyard for intelligent routing in agentic applications; Specific applications: code review, tool use, security monitoring, billing questions; Responds to the growing strength of Chinese open-source models (Alibaba Qwen, Moonshot)
- The model has 30 billion parameters (specific size)
- Released under the Linux Foundation OpenMDW 1.1 license with weights and training recipes
- Accompanying tool NeMo Switchyard for intelligent routing in agentic applications
- Specific applications: code review, tool use, security monitoring, billing questions
- Responds to the growing strength of Chinese open-source models (Alibaba Qwen, Moonshot)
Two audiences, two different impacts
What this means
For individuals
Developers of agentic applications get a smaller, more efficient open-source model (30 billion parameters) that can run on local hardware instead of a paid API, and a complementary tool for routing requests between multiple models.
For a business
Companies that need to run agentic AI locally because of token costs or data sovereignty requirements get a fully open model with weights, code, training recipes and data under the OpenMDW 1.1 license, plus NeMo Switchyard for routing requests across their own mix of models.
DevelopmentCheck the original
Event sources
independently confirmed · 2 publishers, 1 independent. We count feeds from the same owner only once.