NVIDIA: Nemotron 3 Super

What sets this model apart from the obvious hosted assistants is a rare combination: it is completely free through the OpenRouter API, yet ships a 1,000,000-token context window that outsizes most models on this list. The 120B-parameter hybrid MoE design activates only 12B parameters per pass, a Mamba-Transformer architecture aimed at keeping complex multi-agent work efficient. Consider a researcher who needs to load an entire regulatory filing plus a stack of supporting papers into one prompt and then reason across all of it: the wide context plus built-in chain-of-thought handles that without paying per token. Tool and function calling and structured JSON output make it viable as the backbone of an agent that queries APIs and returns parseable results. Unlike Claude Sonnet 5 or ChatGPT, which are polished commercial products with usage costs and broad multimodal support, this trades a paid, image-capable experience for zero cost and open access; note that it is text-only and will not accept image input. It suits builders and researchers who need long-context reasoning and agent scaffolding on a tight or nonexistent budget, and who can live without vision. Choose it for cost-free, large-context text and agent workloads; skip it if you need to feed images or want a supported consumer chat interface.