nemotron-3-super-120b-a12b:freeNVIDIA Nemotron 3 Super is a 120B-parameter open hybrid Mixture-of-Experts model designed for complex multi-agent and long-horizon reasoning workflows. It activates only 12B parameters per token, enabling high compute efficiency while maintaining strong accuracy on advanced tasks. Built on a hybrid Mamba–Transformer MoE architecture with multi-token prediction (MTP), the model delivers significantly higher token generation throughput than leading open models. It supports a 1M-token context window for long-context reasoning, cross-document analysis, and multi-step task planning. Trained with multi-environment reinforcement learning across diverse benchmarks—including AIME 2025, TerminalBench, and SWE-Bench Verified—Nemotron 3 Super achieves strong performance across reasoning and coding tasks. Released fully open with weights, datasets, and training recipes, it supports flexible customization and secure deployment from local workstations to cloud environments.
Select an endpoint and copy a working example for this model.
from openai import OpenAI client = OpenAI( api_key="YOUR_API_KEY", base_url="https://api.apertis.ai/v1") response = client.chat.completions.create( model="nemotron-3-super-120b-a12b:free", messages=[ {"role": "user", "content": "Hello!"} ], max_tokens=1024, temperature=0.7) print(response.choices[0].message.content) # Optional: Enable context compression to reduce token usage# response = client.chat.completions.create(# model="nemotron-3-super-120b-a12b:free",# messages=[{"role": "user", "content": "Hello!"}],# extra_body={"compression": {"enabled": True, "model": "gpt-4.1-mini"}}# )modelmessagesmax_tokenstemperaturetop_pstreamtoolsreasoning_effortstream_optionsthinkingextra_bodyUse these namespaced identifiers in Cursor IDE to avoid conflicts with built-in models.
See how this model compares to others from the same provider.
NVIDIA Nemotron 3 Ultra is an open frontier reasoning and orchestration model featuring a 550B-parameter Mixture-of-Experts (MoE) architecture with 55B active parameters per token. Built on a hybrid Transformer–Mamba design, it supports text input and output with a 1M-token context window, enabling large-scale reasoning and long-horizon task execution. Optimized for agent orchestration, coding agents, deep research, and complex enterprise workflows, the model excels at multi-step reasoning, planning, and sustained execution. With high-throughput inference designed for large-scale agent pipelines, Nemotron 3 Ultra serves as a powerful foundation for advanced agentic AI systems.
NVIDIA Nemotron 3 Ultra is an open frontier reasoning and orchestration model featuring a 550B-parameter Mixture-of-Experts (MoE) architecture with 55B active parameters per token. Built on a hybrid Transformer–Mamba design, it supports text input and output with a 1M-token context window, enabling large-scale reasoning and long-horizon task execution. Optimized for agent orchestration, coding agents, deep research, and complex enterprise workflows, the model excels at multi-step reasoning, planning, and sustained execution. With high-throughput inference designed for large-scale agent pipelines, Nemotron 3 Ultra serves as a powerful foundation for advanced agentic AI systems.
NVIDIA Nemotron 3.5 Content Safety is a compact 4B-parameter multimodal guardrail model from NVIDIA, designed for content moderation, safety classification, and AI policy enforcement. Supporting text and image inputs with text output, it evaluates both user prompts and model responses, providing safe/unsafe classifications, safety category labels, and optional reasoning traces. Fine-tuned from Gemma-3-4B and supporting 12 languages with a 128K-token context window, the model is well suited for prompt moderation, response filtering, content classification, and enterprise safety pipelines. As part of the NVIDIA Nemotron family, it offers a configurable reasoning mode and integrates easily into agentic AI systems requiring robust guardrails and compliance controls.
See how this model compares to others from the same provider.
NVIDIA Nemotron 3 Nano Omni is an open 30B-A3B multimodal model designed as a perception and context sub-agent for enterprise agent systems. It supports text, image, video, and audio inputs with text output, enabling unified multimodal reasoning within a single inference loop. Built on a hybrid MoE Transformer–Mamba architecture with Conv3D video layers and Efficient Video Sampling (EVS), it delivers significantly improved efficiency for video reasoning—achieving ~2× higher throughput and 2.5× lower compute compared to separate pipelines. With up to 300K context length and extended thinking support, it is well suited for scalable, multimodal agent workflows.
NVIDIA Nemotron Nano 2 VL is a 12B open multimodal reasoning model built for video understanding and document intelligence. Using a hybrid Transformer-Mamba design, it delivers high accuracy with lower latency and higher throughput. It handles text and multi-image inputs, excels at OCR, chart and document reasoning, and achieves leading benchmark results across major multimodal tests — including strong performance on long videos via Efficient Video Sampling. Open weights, data, and recipes are available under NVIDIA's permissive license, with broad deployment support across NeMo, NIM, and common runtimes.
NVIDIA Nemotron Nano 2 VL is a 12B open multimodal reasoning model built for video understanding and document intelligence. Using a hybrid Transformer-Mamba design, it delivers high accuracy with lower latency and higher throughput. It handles text and multi-image inputs, excels at OCR, chart and document reasoning, and achieves leading benchmark results across major multimodal tests — including strong performance on long videos via Efficient Video Sampling. Open weights, data, and recipes are available under NVIDIA's permissive license, with broad deployment support across NeMo, NIM, and common runtimes.
Llama-3.3-Nemotron-Super-49B-v1.5 is a 49B reasoning and chat model derived from Llama-3.3-70B-Instruct, tuned for agent workflows like RAG and tool calling with a 128K context window. It combines supervised training with multiple RL stages to improve alignment, step-by-step reasoning, and tool use, while a NAS “Puzzle” architecture reduces memory and boosts throughput so it can run on a single H100/H200. It delivers strong results across math and coding benchmarks, supports toggleable reasoning modes, and is designed for efficient, reliable agent systems and long-context retrieval where accuracy and cost balance matter.
No observed failures in the current observation window