Nemotron-3-Super-120b-a12b: specs, benchmarks, and how to run it on Lyceum

Call Nemotron-3-Super-120b-a12b on Lyceum

To call Nemotron-3-Super-120b-a12b on Lyceum, you only need to update your base URL and API key. Because Lyceum provides an OpenAI-compatible Serverless Inference API, you can use the standard OpenAI Python SDK without rewriting your application logic. This makes it trivial to swap out existing models for NVIDIA's highly efficient LatentMoE architecture. Whether you are building complex multi-agent systems or deploying an autonomous coding assistant, integrating this model takes only seconds.

from openai import OpenAI
 
 client = OpenAI(
 base_url="<the base URL shown in your Lyceum dashboard>",
 api_key="<your lyceum api key>",
 )
 response = client.chat.completions.create(
 model="nvidia/nemotron-3-super-120b-a12b",
 messages=[{"role": "user", "content": "Hello!"}],
 max_tokens=256,
 )
 print(response.choices[0].message.content)

Pricing and region for Nemotron-3-Super-120b-a12b

When you deploy this model through Lyceum's Serverless Inference API, you are billed strictly on a pay-per-token basis. The pricing for Nemotron-3-Super-120b-a12b is set at $0.30 per million input tokens and $0.90 per million output tokens. This model is served through Lyceum Serverless Inference, which prices every model per token with no service tiers and no minimum commitment. This endpoint is globally hosted rather than EU-hosted: it is one of four models in the catalogue that runs outside the EU, and it never receives your traffic unless you explicitly select it. By leveraging Lyceum's infrastructure, you avoid the overhead of provisioning the massive GPU clusters typically required to run a 120B parameter model, paying only for the exact compute you consume during inference. This pay-per-token approach ensures that your infrastructure costs scale linearly with your actual usage, eliminating the financial drain of idle compute instances.

Nemotron-3-Super-120b-a12b use cases

LatentMoE efficiency and architecture

Nemotron-3-Super-120b-a12b is built on a hybrid Latent Mixture-of-Experts (LatentMoE) architecture that interleaves Mamba-2 and MoE layers with select attention layers. While the model contains 120 billion total parameters, it activates only 12 billion parameters per forward pass. This sparse activation allows it to deliver the reasoning capabilities of a massive 120B model while operating with the inference speed and cost profile of a much smaller model. The LatentMoE design routes latent representations rather than raw tokens, allowing the experts to specialize in deep semantic meaning.

Agentic workflows and tool calling

NVIDIA specifically optimized this model for collaborative AI agents and high-volume autonomous workloads, such as IT ticket automation and complex software engineering tasks. It features a configurable reasoning mode that generates an internal thought trace before concluding with a final response. This makes it exceptionally strong at multi-step planning, tool use, and executing agentic loops without losing focus.

128K token context window

Nemotron-3-Super-120b-a12b is served with a 128,000-token context window. This extended memory capacity is critical for Retrieval-Augmented Generation (RAG) pipelines, analyzing entire codebases, or processing extensive document repositories. The hybrid Mamba-Transformer architecture ensures that the model maintains high retrieval accuracy even when the context window is fully saturated, preventing the degradation often seen in standard transformer models.

Nemotron-3-Super-120b-a12b limitations

Where Nemotron-3-Super-120b-a12b falls short

Despite its impressive efficiency, Nemotron-3-Super-120b-a12b comes with specific trade-offs that engineering teams must consider. First, while only 12 billion parameters are active during inference, the model still requires the VRAM to hold all 120 billion parameters in memory. If you were to self-host this model, the BF16 weights alone come to roughly 240 GB, so you would need at least four H100 (80GB) GPUs plus headroom for the KV cache. Using Lyceum's Serverless Inference API abstracts this hardware burden away, but the underlying memory footprint remains massive.

Second, the model's reasoning mode introduces latency and token overhead. When the internal thought trace is enabled, the model generates hidden reasoning tokens before outputting the final answer. While this improves accuracy on complex math and coding problems, it increases the time-to-first-token (TTFT) and inflates the total output token count, which directly impacts your pay-per-token costs. For simple, latency-sensitive tasks like basic text classification or summarization, this overhead is unnecessary, and a smaller dense model would be more cost-effective.

Finally, while the 128,000-token context window is useful, processing prompts of that size requires significant compute time. The Mamba-2 hybrid architecture mitigates some of the quadratic scaling issues of pure attention mechanisms, but saturating the context window will still result in slower response times compared to shorter prompts. Teams should carefully manage their context payloads to balance retrieval depth with application responsiveness.

Nemotron-3-Super-120b-a12b benchmarks

Nemotron-3-Super-120b-a12b benchmark results

NVIDIA designed Nemotron-3-Super-120b-a12b to compete directly with other models in the 120B parameter class, focusing heavily on reasoning, coding, and long-context retrieval. In published evaluations, it demonstrates strong performance across industry-standard benchmarks, particularly excelling in autonomous engineering tasks and mathematical reasoning. Another open model in this size class is covered in our notes on gpt-oss-120b and how to run it.

Below is a comparison of Nemotron-3-Super-120b-a12b against two other models in its weight class: Qwen3.5-122B-A10B and GPT-OSS-120B.

Source: the NVIDIA Nemotron-3-Super-120B-A12B-BF16 model card, Benchmarks table, read 3 August 2026. The Qwen3.5-122B-A10B and GPT-OSS-120B columns are NVIDIA's own published comparison set.

While Qwen3.5-122B-A10B holds a slight edge in general knowledge (MMLU-Pro) and specific coding benchmarks (SWE-Bench), Nemotron-3-Super-120b-a12b maintains a highly competitive profile, particularly in long-context retrieval. Its score of 91.75 on RULER at 1M tokens shows it handles large document payloads without losing fidelity. For teams building agentic workflows that require extensive context, Nemotron offers a highly efficient alternative to dense models. The LatentMoE architecture ensures that even when processing these massive context windows, the inference speed remains exceptionally high compared to traditional dense architectures.

Production deployment for Nemotron-3-Super-120b-a12b

Production configuration for Nemotron-3-Super-120b-a12b

When deploying Nemotron-3-Super-120b-a12b in production, understanding its pricing and configuration parameters is essential for optimizing both performance and cost. As a hybrid MoE model built for multi-agent AI and reasoning, it is served through Lyceum Serverless Inference and billed per token, with no service tiers to choose between. This model is globally hosted rather than EU-hosted.

Because the model supports a 128,000-token context window, you can pass large files or extensive RAG context within that limit. However, you must account for the per-token pricing: $0.30 per million input tokens and $0.90 per million output tokens.

For a realistic production workload, such as an autonomous coding agent analyzing a repository - a single request might consume 50,000 input tokens and generate 2,000 output tokens.

• Input cost: 50,000 tokens * ($0.30 / 1,000,000) = $0.015
• Output cost: 2,000 tokens * ($0.90 / 1,000,000) = $0.0018
• Total cost per request: $0.0168

To maximize the model's reasoning capabilities, you can enable its internal thought trace by passing specific configuration flags in your API request (often handled via extra_body parameters like {"chat_template_kwargs": {"enable_thinking": True}}). When streaming responses in production, ensure your application logic is prepared to handle these reasoning tokens, either by displaying them to the user as a "thinking" state or filtering them out before rendering the final answer.

Why run Nemotron-3-Super-120b-a12b on Lyceum

Why run Nemotron-3-Super-120b-a12b on Lyceum

Lyceum provides a robust, developer-friendly platform for scaling AI workloads without the burden of managing complex hardware. By offering an OpenAI-compatible drop-in replacement API, Lyceum allows engineering teams to switch to Nemotron-3-Super-120b-a12b by updating a base URL and an API key. This eliminates the need to rewrite application logic or learn new SDKs, accelerating your time to market. This model is globally hosted rather than EU-hosted, and it never receives your traffic unless you explicitly select it.

Serverless Inference bills strictly per token, and Lyceum's GPU VMs bill per second when a workload needs dedicated hardware. There are no minimum commitments, no base fees, no idle costs, and zero egress fees, so your spend tracks your real usage instead of reserved capacity. Unified billing gives you a single, predictable view across every model you call, whether you are transitioning off expiring hyperscaler credits or consolidating several providers onto one platform.

You also benefit from our open-stack transparency. Lyceum runs optimized open inference engines like vLLM and NVIDIA Dynamo rather than locking you into proprietary black-box systems, so you can reason about performance, throughput, and cost directly. For background on how this works, see our guide to serverless GPU inference. The combination of drop-in compatibility, transparent open-stack tooling, and pay-per-token economics is what makes Lyceum a strong home for production deployments of Nemotron-3-Super-120b-a12b.

Sources

[1] NVIDIA Nemotron-3-Super-120B-A12B-BF16 Model Card; [2] NVIDIA NeMo Evaluator (GitHub); [3] NVIDIA NeMo Evaluator: Nemotron 3 Super reproducibility guide; [4] Ollama Library: Nemotron 3 Super