NVIDIA: Nemotron 3 Super (free)

NVIDIA Nemotron 3 Super is a text model from NVIDIA, released on 11 March 2026. It is a 120B-parameter open hybrid mixture-of-experts model that activates just 12B parameters, which NVIDIA describes as a way to get compute efficiency and accuracy in complex multi-agent applications. It is built on a hybrid Mamba-Transformer architecture. On api14 it is listed under the model id nvidia/nemotron-3-super-120b-a12b:free.

The model works with text input and produces text output. Its context window is 262,144 tokens, and the maximum output is 235,929 tokens, so it can hold very long inputs and still return long generations in a single request. Because it is a reasoning-capable model, responses can include a reasoning component, and the supported parameters include include_reasoning, reasoning, and reasoning_effort for controlling that behaviour.

Those capabilities point to use cases where long context and multi-step reasoning matter. A 262k-token window suits work over large documents, long conversation histories, or codebases that would not fit in smaller models. The reasoning controls and the multi-agent framing in NVIDIA's description make it a reasonable candidate for agent-style workflows, tool-driven pipelines, and tasks that need structured output. The large maximum output also helps when a single answer has to be long, such as a detailed report or an extended extraction over many records.

You call it through api14's OpenAI-compatible API using the model id nvidia/nemotron-3-super-120b-a12b:free. The request parameters it supports are include_reasoning, max_tokens, reasoning, reasoning_effort, response_format, seed, structured_outputs, temperature, tool_choice, tools, and top_p. In practice that means you can set sampling with temperature and top_p, cap the reply with max_tokens, make runs reproducible with seed, request JSON or schema-constrained output with response_format and structured_outputs, and give the model callable functions through tools and tool_choice. The reasoning parameters let you decide whether reasoning is returned and how much effort is spent on it.

Two limits are worth noting before you commit. Input is text only, so this model is not the right choice if you need to send images or audio. It is also a large model, and the reasoning and tool features add request complexity, so for short, simple, latency-sensitive text tasks a smaller model may fit better.