Nemotron-4

Nemotron-4 is NVIDIA’s fourth-generation family of large language models released as open weights for commercial use. It comes in multiple sizes—including very large (e.g., ~340B) and smaller, cost-efficient variants—each available as base (pretrained) and instruct (chat-aligned) checkpoints. The instruct models are tuned for step-by-step reasoning, code generation/refactoring, and reliable instruction following; the base models are ideal starting points for domain fine-tuning or custom alignment.

Key capabilities

Long-context prompting (repo-/document-scale), streaming, and deterministic JSON modes

Function/tool calling for agents and RAG (retrieval, search, code exec, SQL)

Strong coding skills across common languages and solid math/analysis performance

Synthetic-data generation workflows to bootstrap or improve smaller task models

Engineering & deployment

Optimized inference with TensorRT-LLM and Triton; packaged as NIM microservices for easy, scalable hosting

Quantization options (8/4-bit) and multi-GPU parallelism for low latency and high throughput

Plays well with popular runtimes (vLLM/Transformers) and guardrail stacks (e.g., policy filters)

Typical uses

Enterprise copilots over private data (RAG)

Coding assistants, analytics/BI agents, and structured-output automations

Programmatic data generation for fine-tuning other models

Choose Nemotron-4 Instruct when you need a ready-to-use, safe-by-default assistant; pick Nemotron-4 Base when you want maximum control for custom adaptation.