NVIDIA: Nemotron 3 Super

Nemotron 3 Super is an NVIDIA text model, published under the model id nvidia/nemotron-3-super-120b-a12b. NVIDIA describes it as a 120B-parameter open hybrid MoE model that activates just 12B parameters, built on a hybrid Mamba-Transformer design. The "Super" tier sits within NVIDIA's Nemotron 3 family, and the provider release date is 2026-03-11. It is a text-only model, so it does not accept images or audio, and it produces text output.

The model accepts text input and returns text output. Its context window is 262,144 tokens, and the maximum output is 235,929 tokens, so long inputs and long generations are both possible within a single request. The supported parameters include the usual sampling controls (temperature, top_p, top_k, min_p, repetition_penalty, frequency_penalty, presence_penalty, seed, stop, logit_bias) as well as reasoning controls (reasoning, reasoning_effort, include_reasoning). It also supports logprobs and top_logprobs, response_format, structured_outputs, tools and tool_choice.

Those capabilities point to a few practical uses. The large context window suits work that involves long documents, such as summarising or extracting structured data from lengthy reports, contracts or transcripts. Tool calling and structured outputs make it a candidate for agent-style workflows where the model has to decide which function to call and return arguments in a fixed schema. The reasoning parameters let you trade off how much internal reasoning the model does, which is useful for tasks that need multi-step analysis rather than quick replies. Because it is text-only, it fits pipelines that are already text based and do not need vision.

You call it through api14's OpenAI-compatible API by setting the model field to "nvidia/nemotron-3-super-120b-a12b". The request body follows the OpenAI chat completions shape, so messages, tools, response_format and the sampling parameters listed above can be passed directly. If you want the model's reasoning exposed, set include_reasoning or use the reasoning and reasoning_effort parameters. For deterministic or reproducible runs, use seed together with temperature. If you need JSON output, response_format or structured_outputs are available.

There are a couple of things to keep in mind. The model is text only, so if your application needs to read images or audio, a multimodal model is a better fit. The context window is large but finite at 262,144 tokens, and the output cap is 235,929 tokens, so very large batches of documents may still need chunking. Reasoning modes can add latency, so for simple classification or short replies you may prefer a smaller or non-reasoning model.