NVIDIA: Nemotron 3 Ultra (free)

NVIDIA Nemotron 3 Ultra is an open frontier reasoning and orchestration model from NVIDIA. It is a mixture-of-experts model with 55B active parameters out of 550B total, built on a hybrid Transformer-Mamba architecture. The provider release date is 2026-06-04. On api14 it appears as a text model under the identifier nvidia/nemotron-3-ultra-550b-a55b:free, and the free suffix marks the variant available on this gateway.

The model accepts text input and returns text output, so it fits text-only pipelines rather than multimodal work. Its context window is 1,000,000 tokens, which is large enough to hold entire codebases, long document sets or extended conversation histories in a single request. The maximum output is 65,536 tokens, so responses can be long, but they are still bounded well below the input capacity. Because it is a reasoning model, it can expose its reasoning process, and the supported parameters include reasoning, reasoning_effort and include_reasoning alongside the usual sampling controls.

Those capabilities point to a few practical uses. The long context suits analysis over large repositories, legal or technical document collections, and multi-file refactors where the model needs to see many parts of a project at once. The reasoning and orchestration focus makes it a candidate for agent-style workflows, planning steps and multi-stage tasks. Tool calling support means it can be wired into function-calling loops where the model decides which tool to invoke, and the seed parameter helps when you want repeatable sampling during testing.

You call it through api14's OpenAI-compatible API using the model id nvidia/nemotron-3-ultra-550b-a55b:free. The supported request parameters are include_reasoning, max_tokens, reasoning, reasoning_effort, seed, temperature, tool_choice, tools and top_p. In practice you send a chat completions request with messages, set max_tokens to cap the 65,536-token output ceiling, adjust temperature and top_p for sampling, and pass tools plus tool_choice when you want function calling. Use reasoning_effort or reasoning to control how much reasoning the model produces, and include_reasoning if you want the reasoning returned in the response.

Keep the 1,000,000-token context and 65,536-token output limits in mind when sizing requests, since prompts and outputs share the same budget in a conversation. If your task is purely text generation at small scale, a lighter model may be simpler, but for long-context reasoning, orchestration and tool use, this model covers those needs directly.