After Meta, Nvidia releases 30B open-weight AI model for agentic tasks

Nvidia has introduced Nemotron 3.5 Lightning, a 30-billion-parameter open model for agentic workloads, as Nvidia, Meta and others back a broader industry push for open-weight AI

Nvidia has introduced a new 30-billion-parameter open AI model called Nemotron 3.5 Lightning for agentic workloads, a day after Meta released its own 30-billion-parameter open-weight model, Muse Glimmer, as both companies back the growing industry push for open AI models. Nvidia CEO Jensen Huang, along with Meta and other technology companies, recently formed a coalition supporting open-weight AI models that has urged policymakers to avoid broad restrictions on models whose weights can be downloaded and modified. 

Nvidia Nemotron 3.5 Lightning: Details

Called Nemotron 3.5 Lightning, Nvidia's model is designed for the repetitive execution tasks performed by AI agents, such as making tool calls, validating results and delegating work to subagents. The company said these tasks can account for much of the workload of long-running agents, making the use of larger reasoning models at every step costly and slow.

 

 

Today, NVIDIA announced NVIDIA Nemotron 3.5 Lightning, a customizable model for high-volume, specialized work, and NVIDIA NeMo Switchyard, which helps agents route each workflow step across the models they choose. ⚡ https://t.co/Li96xrOe3K

— NVIDIA (@nvidia) August 11, 2026

Unlike a conventional 30-billion-parameter model, Nemotron 3.5 Lightning uses a mixture-of-experts architecture with only 3 billion parameters active for each token. Nvidia said this allows the model to handle the capacity of a larger model while reducing the amount of computation required for each request.

 

Nvidia said Nemotron 3.5 Lightning is intended to work alongside larger reasoning models. While models such as Nemotron 3 Ultra can handle planning and orchestration, Lightning is designed for high-volume execution once an agent has decided what needs to be done.

The model is also trained for agent harnesses. It supports agent frameworks including OpenClaw and Hermes Agent, which are also supported by Nvidia's NemoClaw security and management stack.

 

Nvidia is also releasing NeMo Switchyard, a model-routing library that can direct different tasks to different models. Under this approach, complex planning can be routed to a larger model while routine execution is handled by Nemotron 3.5 Lightning.

Nvidia claims faster execution

According to Nvidia, Nemotron 3.5 Lightning reached 86 per cent accuracy on PinchBench while completing 10,000 tasks 30 per cent faster than Qwen3.6 35B at a similar level of accuracy. The company said the model can deliver up to four times the output speed of similarly sized models.

 

The model uses speculative decoding, which allows a draft model to propose multiple tokens that can then be reviewed by the main model. Nvidia said it also ships with two draft models, DSpark and DFlash, for different inference workloads.

Open weights and local deployment

Nvidia is releasing the model with its weights, training data and recipes under the OpenMDW-1.1 licence. Developers can fine-tune it using Nvidia's NeMo tools and use reinforcement learning and environment-based evaluations.

 

Nemotron 3.5 Lightning can also be deployed locally on systems including Nvidia's DGX Spark, Jetson devices and GeForce RTX 5090. Nvidia said it works with tools including LM Studio, llama.cpp, Ollama and Unsloth.

 

The release comes as Nvidia and other technology companies continue to push the case for open-weight AI models. A letter backed by Nvidia, Meta, Microsoft, OpenAI, Google and other companies argued that open-weight models should not face broad restrictions, saying they can support competition, innovation and AI development.

 

Meta's release of Muse Glimmer a day earlier and Nvidia's Nemotron 3.5 Lightning now put two 30-billion-parameter open models from major technology companies in the spotlight.