Nemotron

Nemotron is a family of open foundation models developed by NVIDIA for language, vision-language and agentic workloads. Unlike NVIDIA's better-known role as a chipmaker, Nemotron places the company directly in the model layer: the family is distributed with open weights and, in many releases, with training data recipes and reward models, aimed at enterprises that want to customise and self-host frontier-class models.^[1]

History

NVIDIA introduced the Nemotron line in 2024 with Nemotron-4 (15B and 340B parameters), released alongside Nemotron-4 340B Reward — a set of reward models used to filter and rank training data, which NVIDIA published so others could reproduce its data-curation pipeline. The Llama 3.1 Nemotron series followed, offering instruction-tuned and reward-model variants built on Meta's open Llama weights.^[1]

The family expanded through 2025 into reasoning and hybrid architectures (Nemotron-H) and, in 2026, into a broad Nemotron 3.5 generation. On 15 October 2026 NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter mixture-of-experts vision-language model positioned for always-on agents, together with edge variants Nemotron 3.5 Bolt 74B-A9B (tuned to fit 24 GB of memory) and Nemotron 3.5 Bolt E4B (a 4 GB model for embedded deployment). The earlier Nemotron 3.5 Lightning, a 30B-total/3B-active mixture-of-experts model, was made freely available through model routers.^[2][3]

Key Concepts and Technology

Three traits define the family. Open weights — Nemotron models can be downloaded and run on-premises, a requirement for Malaysian banks and public-sector bodies that cannot send data to external APIs. Agentic optimisation — the 3.5 generation is benchmarked specifically for task completion by persistent, always-on agents rather than one-shot chat. Edge scaling — the Bolt variants trade a fraction of intelligence for the ability to run on a single workstation GPU or local device, supporting NVIDIA's strategy of scaling models from data centre to desktop.^[2]

Nemotron releases are deeply integrated with NVIDIA's software stack (NeMo for training and evaluation, TensorRT-LLM and NIM for inference), so the family also functions as a reference workload demonstrating the value of NVIDIA hardware.^[3]

Applications and Impact

Nemotron is used for building custom enterprise assistants, retrieval pipelines, vision-language inspection systems and on-device agents. NVIDIA's open data and reward-model releases have been particularly influential: the Nemotron reward models became a standard tool in the community for synthesising preference data for RLHF without collecting new human annotations.^[1] By giving away models, NVIDIA lowers the cost of adopting the very workloads that drive demand for its accelerators — an explicit element of the company's ecosystem strategy.

>See Also

• NVIDIA
• Open Weight Models
• Mixture of Experts
• NVIDIA DGX
• Sovereign AI
• Edge AI