NVIDIA released Nemotron 3 Super at GTC 2026 — a hybrid Mamba-Transformer model with the highest SWE-Bench Verified score of any open-weight model (60.47%) and 2.2x the throughput of GPT-OSS-120B. Here is what developers need to know.
On March 11, 2026, during NVIDIA's GTC conference, the company released something that quietly rewrote the leaderboard for open AI models. NVIDIA Nemotron 3 Super is a 120-billion-parameter hybrid model that scores 60.47% on SWE-Bench Verified — the most rigorous coding benchmark in AI — beating GPT-OSS-120B's 41.90% by nearly 20 points. It also delivers 2.2 times the inference throughput at the same time.
For developers and enterprises that have been waiting for an open-weight model that genuinely competes with closed frontier models on real-world coding tasks, this is that announcement.
What Is NVIDIA Nemotron 3 Super?
Nemotron 3 Super is NVIDIA's first open model built specifically for agentic AI workloads — the kind of multi-step, multi-tool tasks where an AI needs to maintain context across an entire codebase, reason over long documents, and take sequences of actions to complete complex goals.
The core specs:
- 120 billion total parameters, 12 billion active per token (MoE architecture)
- 1 million token context window — enough to fit an entire mid-size codebase in a single prompt
- Training across 10+ reinforcement learning environments
- Support for 20 languages and 43 programming languages
- A novel LatentMoE architecture that activates 4x more experts at the same computational cost
The model is available for free on OpenRouter and Hugging Face, and for enterprise deployment through NVIDIA NIM containers.
The Architecture That Makes It Different
Nemotron 3 Super is NVIDIA's first model to combine three distinct architectural paradigms into a single system, each addressing a different limitation of existing approaches.
Mamba-2 State Space Layers
The majority of sequence processing is handled by Mamba-2 layers — state space models (SSMs) that offer linear-time complexity with respect to sequence length. Traditional attention layers scale quadratically as sequences get longer, making long-context reasoning computationally expensive. Mamba-2's linear scaling is what makes the 1M token context window practical and fast, not just theoretically possible.
Standard Transformer Layers
Interspersed throughout are standard transformer layers for attention-based reasoning. The combination — SSMs for efficient sequence processing, attention for precise reasoning — gives Nemotron 3 Super the best properties of both architectures without the worst costs of either.
LatentMoE: NVIDIA's Novel Expert Routing
This is NVIDIA's most significant architectural innovation in Nemotron 3 Super. LatentMoE compresses input tokens into a latent space before routing them to experts. This compression allows the system to activate four times more experts at the same computational cost as traditional MoE routing. More experts per token means more specialized knowledge brought to bear on each computation — without a corresponding compute increase.
What this means in practice: Nemotron 3 Super gets more inputs from specialized sub-networks on each computation, resulting in richer, more accurate outputs on complex tasks — especially coding and multi-step reasoning — without making the model slower or more expensive to run.
Multi-Token Prediction (MTP)
Nemotron 3 Super uses Multi-Token Prediction for speculative decoding. Instead of predicting one token at a time, MTP predicts multiple future tokens simultaneously and verifies them in one pass. On SPEED-Bench, this achieves an average acceptance length of 3.45 tokens per verification step compared to 2.70 for DeepSeek-R1, translating to up to 3x wall-clock speedup without needing a separate draft model.
The Benchmark Numbers
Benchmarks have become almost meaningless in a world where every model claims to beat every other model. But a few specific results for Nemotron 3 Super are genuinely hard to dismiss.
SWE-Bench Verified: #1 Open-Weight Model
SWE-Bench is arguably the most meaningful coding benchmark available. Unlike question-answering tests, it measures whether a model can resolve real GitHub issues — reading code, understanding bugs, writing fixes, and passing automated tests. These are the exact tasks that matter in production agentic coding systems.
Nemotron 3 Super scores 60.47% on SWE-Bench Verified. Compare that to:
- GPT-OSS-120B: 41.90%
- Qwen3.5-122B: ~51%
- DeepSeek-R1: ~49%
A gap of nearly 20 percentage points over GPT-OSS is significant on a benchmark this rigorous. This isn't a rounding error or a prompt engineering trick — it reflects a real difference in how reliably the model navigates real-world code at scale.
Long-Context Retention: RULER at 1M Tokens
Most models claim a large context window but lose coherence long before they reach it. RULER measures how well a model actually retains and uses information at different context lengths. At 1 million tokens, Nemotron 3 Super scores 91.75% on RULER. GPT-OSS-120B scores 22.30% at the same length.
This is the difference between a model that technically accepts a million tokens and one that actually understands the content at that scale. For agentic coding systems that need to read and reason over entire repositories, this distinction is everything.
DeepResearch Bench: #1 Overall
Nemotron 3 Super holds the #1 position on the DeepResearch Bench, which measures an AI's ability to conduct thorough, multi-step research across large document sets — finding relevant information, synthesizing it across sources, and answering complex questions that require reading and reasoning simultaneously.
Inference Throughput
This is where Nemotron 3 Super's architecture pays the clearest dividends at production scale:
- 2.2x higher throughput than GPT-OSS-120B
- 7.5x higher throughput than Qwen3.5-122B
At production scale — millions of requests per day — throughput is money. Higher throughput means more requests processed on the same hardware, directly reducing per-request cost. For companies building on open models, this throughput advantage could translate to 50-80% infrastructure cost reduction compared to alternatives with similar accuracy.
How to Access Nemotron 3 Super
Free Access via OpenRouter
The fastest path to testing Nemotron 3 Super is OpenRouter, which offers the model at no cost with rate limits. This is ideal for experimentation, evaluation, and small-scale use cases. No infrastructure required, no NVIDIA account needed — try it from your browser today.
Hugging Face (Self-Hosted)
The model weights are on Hugging Face under NVIDIA's Open Model License Agreement. Download and run using vLLM, TensorRT-LLM, or other inference frameworks for full control over your deployment. Note that NVIDIA's license is not Apache 2.0 — review the terms before building commercial products on top of it.
NVIDIA NIM (Enterprise)
For production enterprise deployments, NVIDIA offers Nemotron 3 Super as a NIM container — fully optimized for NVIDIA GPU infrastructure with enterprise SLAs, support contracts, and performance guarantees. This is the path for organizations that need reliability at scale.
build.nvidia.com
NVIDIA's API playground lets you test the model in the browser before committing to an integration. The full 1M context window is available for testing, including the long-context analysis that makes this model compelling.
Real-World Integrations Already Shipping
Early adoption of Nemotron 3 Super is concentrated in agentic coding tools — exactly the use case NVIDIA designed it for:
- CodeRabbit: AI code review tool using Nemotron 3 Super for deeper codebase analysis and more accurate pull request reviews
- Factory: Agentic software development platform integrating Nemotron 3 Super for multi-step coding tasks
- Greptile: Codebase search and understanding tool leveraging the model's long-context capabilities to analyze large repositories
These aren't experiments — these are commercial products where teams have evaluated alternatives and selected Nemotron 3 Super for production coding workloads. The SWE-Bench number gets confirmed in practice.
The Licensing Reality
Nemotron 3 Super is released under the NVIDIA Open Model License Agreement (updated October 2025). This is more permissive than most enterprise licenses, but it is not Apache 2.0 or MIT. The license includes safeguard clauses that restrict certain high-risk applications.
For regulated industries or use cases where licensing certainty is critical, review the license terms carefully. The safeguard clauses are designed to prevent misuse, but they add legal review complexity that Apache 2.0 models like Mistral Small 4 don't require. For many enterprise use cases this is fine — for others it's a meaningful consideration.
Who Should Use Nemotron 3 Super?
Enterprise Coding Teams
Any team building or using agentic coding tools — code review, automated PR analysis, multi-file refactoring, bug detection — should evaluate Nemotron 3 Super. The SWE-Bench lead over every other open-weight model is the clearest available signal for coding performance. This is the model to benchmark against for any coding-focused AI deployment in 2026.
Teams Working With Large Codebases
The 1M token context with 91.75% RULER retention means Nemotron 3 Super can process entire medium-sized codebases without truncation — and actually understands them. For organizations that need AI to navigate their full codebase rather than fragments of it, this is a meaningful capability that alternatives don't match.
Cost-Sensitive AI Builders
If you're currently using a closed-source model for coding or research tasks and paying commercial API rates, Nemotron 3 Super's throughput efficiency could dramatically reduce infrastructure costs in a self-hosted deployment. The combination of benchmark leadership and throughput advantage makes the economics compelling at scale.
Multi-Agent System Builders
NVIDIA designed Nemotron 3 Super explicitly for multi-agent architectures. Its training across 10+ reinforcement learning environments makes it more reliable as an autonomous agent than models primarily trained on supervised data. If you're building systems where AI agents need to plan, execute, and self-correct across long task horizons, this model's design directly addresses your requirements.
Limitations to Know
- License restrictions: Not Apache 2.0. Review NVIDIA's Open Model License before commercial deployment, especially for regulated industries.
- Hardware requirements: 120B total parameters require significant GPU infrastructure for self-hosting. Budget at minimum 4-8 high-end GPUs for production throughput.
- Agentic-first design: Nemotron 3 Super is optimized for coding and long-context agentic workloads, not creative writing or general consumer tasks. For those use cases, other models may be more appropriate.
- Newer ecosystem: As a March 2026 release, community tooling, fine-tuning recipes, and unofficial deployment guides are still maturing compared to older models with larger ecosystems.
People Also Ask
Is NVIDIA Nemotron 3 Super really free?
Yes, the model is available for free through OpenRouter with rate limits, and the weights are available on Hugging Face under NVIDIA's Open Model License. Enterprise access through NVIDIA NIM is available at commercial rates with full support. The OpenRouter free tier is sufficient for most experimentation and evaluation.
How does Nemotron 3 Super compare to DeepSeek R1?
Nemotron 3 Super leads on SWE-Bench Verified (60.47% vs DeepSeek-R1's ~49%) and RULER long-context retention (91.75% at 1M tokens vs DeepSeek-R1's lower scores at that length). DeepSeek-R1 has strong math and reasoning profiles. For coding and long-context agentic tasks specifically, Nemotron 3 Super's combination of SWE-Bench performance and context retention is currently superior among open-weight models.
What is LatentMoE?
LatentMoE is NVIDIA's novel expert routing architecture in Nemotron 3 Super. It compresses tokens into a latent space before routing them to experts. This compression allows the system to activate four times more experts at the same computational cost as standard MoE routing, improving output quality without increasing inference expense.
Can Nemotron 3 Super run locally?
With appropriate hardware, yes. The model's MoE design means only 12B parameters are active per token, which is more feasible than a dense 120B model. However, storing the full 120B parameter set still requires substantial GPU memory. For most teams, NVIDIA NIM or the free OpenRouter access will be more practical than local self-hosting.
Want to skip months of trial and error? We've distilled thousands of hours of prompt engineering into ready-to-use prompt packs that deliver results on day one. Our packs at wowhow.cloud include battle-tested prompts for marketing, coding, business, writing, and more — each one refined until it consistently produces professional-grade output.
Blog reader exclusive: Use code
BLOGREADER20for 20% off your entire cart. No minimum, no catch.
Written by
Promptium Team
Expert contributor at WOWHOW. Writing about AI, development, automation, and building products that ship.
Ready to ship faster?
Browse our catalog of 1,800+ premium dev tools, prompt packs, and templates.