Nemotron 3 Super 120B A12B
Neural Network
Nemotron 3 Super 120B A12B is an open hybrid MoE model by NVIDIA for complex multi-agent systems and workflows
Specifications
- Input
- text, images, documents
- Output
- text
- Context size
- 1 048 576 tokens
- Max answer length
- 262 144 tokens
- Reasoning effort
- none, low, medium
- Capabilities
- reasoning, function calling
Prompt cost
(per 1M tokens)
Answer cost
(per 1M tokens)
How it works Nemotron 3 Super 120B A12B?
Nemotron 3 Super 120B A12B is an open NVIDIA model with a hybrid Mamba and Transformer architecture. Out of 120 billion parameters, about 12 billion are activated per request, so the Mixture of Experts (MoE) scheme provides accuracy with efficient computation. Developers targeted complex multi-agent applications where multiple roles exchange replies and tools. In BotHub, the model is accessible from Russia without a VPN or foreign card: you pay in rubles per token, not for a subscription, and tokens do not expire. Over 250 models are available in one window, making it easy to switch and compare answers. A unified OpenAI-compatible API allows replacing the model in a multi-model system without rewriting the integration. For companies, there is a contract, invoice, and admin panel with limits. It is useful when building an agent chain with external service calls, automating document parsing and analytics, building a RAG assistant over your knowledge base, prototyping internal team tools, or looking for a more computationally efficient option for high-load environments.Frequently asked questions about Nemotron 3 Super 120B A12B
You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.
This is an NVIDIA text model with a reasoning mode. It is suitable for working with code, analyzing and parsing documents, STEM tasks and logic, extracting data from images, as well as for assistants and pipelines that require calling external tools. Context is 8000 tokens.
The output limit is very large — up to 235,929 tokens, so the model can output long analyses, large code modules, and detailed documents in full. However, the context is 8000 tokens, so it is wiser to provide large input materials in parts. You pay per token, without a subscription.
Yes, reasoning before answering is confirmed by our data: the model first breaks down the task into steps and only then formulates a conclusion. This significantly helps with math, logical chains, code debugging, and multi-step scenarios. This mode is usually not needed for simple queries.
Yes, function calling and structured JSON output are supported. The model can be connected to external services, databases, and custom tools, receiving responses strictly according to the schema. BotHub uses a unified OpenAI-compatible API, so replacing the model does not require rewriting your integration.
Images and documents are accepted as input: ask to describe a picture, analyze a diagram or table, or extract data from a scan. Audio and video are not noted in our data, and image generation is not claimed — the model outputs text.
We do not have specific data on the quality of Russian, so it is more reliable to test the model with your typical queries and prompts. BotHub has over 250 models in one window: results are easy to compare and switch between, without VPN or foreign cards, with payment in rubles.