Summary: NVIDIA released Nemotron-3-Super-120B-A12B under its open model license on March 11, 2026, with 120 billion total parameters, 12 billion active parameters, and a one-million-token context window. NVIDIA opened weights, datasets, RL environments, evaluation recipes, and a technical report; current channels include Hugging Face and OpenRouter, while AWS Bedrock and Azure are coming soon. Independent tests put intelligence behind frontier models but logged 478 output tokens per second as Artificial Analysis’s fastest measurement, fitting the release’s throughput-versus-intelligence trade-off.
Nemotron-3-Super-120B-A12B went out under NVIDIA’s open model license on March 11, 2026: 120 billion parameters in total, 12 billion of them active on any given pass, and a context window of one million tokens. Chasing the “smartest model” crown was never the plan. Independent benchmarks place its composite intelligence score a long way behind Gemini 3.1 Pro and GPT-5.4. What the model does own is a speed record — 478 output tokens per second, the highest reading Artificial Analysis has ever logged. NVIDIA’s wager is easy to spell out: in the agent era, what caps you is seldom how sharp one answer is; it is the inference bill that piles up over long, multi-step, tool-calling runs.
The schedule was chosen with care. Second in the Nemotron 3 family, it follows December’s Nemotron 3 Nano (30B) by a mere 86 days, and the reveal landed in the week leading up to GTC.
Out-Opening the Rivals Just Before GTC
One detail The New Stack flagged says a lot: the roadmap originally called for 100B total and 10B active parameters, yet the shipping model came in at 120B/12B — dimensions that set it directly against OpenAI’s gpt-oss-120B. The open-model market in 2026 has become a straight-up fight, a divergence we traced in our opening-of-year outlook. NVIDIA’s counter is to open wider than anyone: the weights, yes, but the datasets and training environments come along as well.
The Architecture Play: Mamba Paired with Latent MoE
Its edge comes from how the model is built:
- Latent MoE fires up four experts at a time while charging for one
- Hybrid Mamba-Transformer layers: per NVIDIA, memory and compute efficiency sit 4x above a pure transformer stack, which helps when long tasks must keep context in view
- Multi-Token Prediction: credited by NVIDIA with a 3x inference speedup
- A 1M-token window that can hold a large codebase in a single load
Every one of these decisions answers the same pain point: agents and their context explosion. NVIDIA reckons a single agent step can put out as much as 15x the tokens of a plain chat exchange, and the “thinking tax” reasoning models levy keeps compounding across turns of tool calls.
Vendor Efficiency Claims Against Independent Scores
The numbers NVIDIA publishes are bold. Versus the previous-generation Nemotron Super, throughput climbs by up to 5x and accuracy by 2x. On Blackwell silicon running NVFP4, the model is up to 4x quicker than FP8 on Hopper while losing no accuracy. NVIDIA further claims a 7.5x inference-throughput edge over Qwen3.5-122B, and says the model carried NVIDIA AI-Q to first place on both DeepResearch Bench and DeepResearch Bench II.
Independent results read cooler. Artificial Analysis assigns an overall intelligence score of 36 — a shade above gpt-oss-120B at 33, far beneath the 57 posted by Gemini 3.1 Pro and GPT-5.4, and lower than Qwen3.5-122B, DeepSeek V3.2, and GLM-5. The speed claim survives scrutiny: 478 tokens of output per second is the fastest the organization has ever logged, and gpt-oss-120B trails in second at 264. Intelligence was knowingly traded away for throughput — and once a task demands hundreds of tool calls, that trade starts to look exactly right.
What Ships in the Open Release
Measured against most rivals, NVIDIA opened a great deal:
- The model weights, released under the NVIDIA Nemotron Open Model License
- Pre-training and post-training datasets exceeding 10 trillion tokens, synthetic portions of which were generated by frontier reasoning models
- Fifteen RL training environments, evaluation recipes, and a complete technical report
- An end-to-end fine-tuning path through NVIDIA NeMo
Picking the model up is simple. Pull it straight from build.nvidia.com, Hugging Face, OpenRouter (free at present), or Perplexity; enterprise routes run through Google Vertex AI and Oracle Cloud, with AWS Bedrock and Azure flagged as coming soon. A NIM microservice packaging is included too, for self-hosting on CoreWeave, Crusoe, Nebius, Together AI, and other platforms.
How Developers and Platform Teams Should Read It
One: the pairing of “smart enough, extremely fast, and self-hostable” meets genuine demand. Perplexity, CodeRabbit, Factory, Greptile, Palantir, Cadence, Dassault Systèmes, and Siemens are already using it, with workloads clustered around whole-codebase loading, long-document financial analysis, and tool calling that has to hold up. Two: because datasets and RL environments ship alongside, fine-tuning teams can retrace the original training path instead of working from final weights alone — a real difference for anyone pursuing domain specialization. Three: a bigger Nemotron 3 Ultra (550B total parameters) is still to come. If it keeps the same trade-off philosophy, the open camp will go on judging itself against the closed frontier by efficiency, not by leaderboard placement.