Nemotron 3 Super (free)

Kilo Code documents the free Nemotron 3 Super variant as a 120B-parameter hybrid MoE model activating 12B parameters, with a 262,144-token context window, 235,929 max output tokens, text input, and no input or output pricing on its coding-agent platform. It reports 15.5% completion on Terminal Bench 2.0 at zero cost per attempt from official Kilo evaluations. Community usage peaked at roughly 22 million weekly tokens around late September, with debugging mode showing a score of 78 on Kilo's leaderboard and weekly token volume ranging from about 5.5M to 22M. Function calling, tool choice, structured JSON outputs, and reasoning tokens are all supported, and the model was created March 11, 2026 with content moderation disabled.

The OpenRouter compare page for Nemotron 3 Super (free) wires the model into comparison views against flagship models such as Claude Fable 5 (batch), GPT-5.5 (batch), and Gemini 3.1 Pro Preview (batch), as well as affordable picks including DeepSeek V4 Flash 0423, Hy3 preview, and Gemini 2.5 Flash Lite (batch). It also Technically, the comparison landing page offers no new specs beyond the model page and is mostly scaffolding for side-by-side benchmarking, pricing, and context-length comparisons. The excerpt is heavily UI chrome and competitive-group labels, so fresh technical signal is limited, but it confirms Nemotron 3 Super (free

NVIDIA Nemotron 3 Super is a 120B-parameter open hybrid Mamba-Transformer Mixture-of-Experts model that activates only 12B parameters, with Multi-Token Prediction enabling over 50% higher token generation than leading open models. Latent MoE lets it call four experts at the cost of one, and multi-environment RL training across 10+ environments targets accuracy on AIME 2025, TerminalBench, and SWE-Bench Verified. The free variant ships with a 262K token context window on OpenRouter and reports Artificial Analysis scores of 12.8 Intelligence Index, 37.7 Coding Index, 1.7 Agentic Index, and 80.0% on GPQA Diamond under its reasoning configuration. Weights, datasets, and recipes are released under the NVIDIA Open License, allowing customization from workstation to cloud.

Effective pricing across providers for NVIDIA: Nemotron 3 Super (free) - NVIDIA Nemotron 3 Super is a 120B-parameter open hybrid MoE model, activating just 12B parameters for maximum compute efficiency and accuracy in complex multi-agent applications. Built on a hybrid Mamba-Transformer Mixture-of-Experts architecture