Launched as we speak, NVIDIA Nemotron 3 Tremendous is a 120‑billion‑parameter open mannequin with 12 billion lively parameters designed to run complicated agentic AI methods at scale.
Out there now, the mannequin combines superior reasoning capabilities to effectively full duties with excessive accuracy for autonomous brokers.
AI-Native Corporations: Perplexity provides its customers entry to Nemotron 3 Tremendous for search and as one among 20 orchestrated fashions in Pc. Corporations providing software program improvement brokers like CodeRabbit, Manufacturing unit and Greptile are integrating the mannequin into their AI brokers together with proprietary fashions to attain increased accuracy at decrease value. And life sciences and frontier AI organizations like Edison Scientific and Lila Sciences will energy their brokers for deep literature search, knowledge science and molecular understanding.
Enterprise Software program Platforms: Business leaders corresponding to Amdocs, Palantir, Cadence, Dassault Systèmes and Siemens are deploying and customizing the mannequin to automate workflows in telecom, cybersecurity, semiconductor design and manufacturing.
As firms transfer past chatbots and into multi‑agent purposes, they encounter two constraints.
The primary is context explosion. Multi‑agent workflows generate as much as 15x extra tokens than commonplace chat as a result of every interplay requires resending full histories, together with device outputs and intermediate reasoning.
Over lengthy duties, this quantity of context will increase prices and might result in purpose drift, the place brokers lose alignment with the unique goal.
The second is the considering tax. Complicated brokers should cause at each step, however utilizing massive fashions for each subtask makes multi-agent purposes too costly and sluggish for sensible purposes.
Nemotron 3 Tremendous has a 1‑million‑token context window, permitting brokers to retain full workflow state in reminiscence and stopping purpose drift.
Nemotron 3 Tremendous has set new requirements, claiming the highest spot on Synthetic Evaluation for effectivity and openness with main accuracy amongst fashions of the identical dimension.
The mannequin additionally powers the NVIDIA AI-Q analysis agent to the No. 1 place on DeepResearch Bench and DeepResearch Bench II leaderboards, benchmarks that measure an AI system’s potential to conduct thorough, multistep analysis throughout massive doc units whereas sustaining reasoning coherence.
Hybrid Structure
Nemotron 3 Tremendous makes use of a hybrid combination‑of‑consultants (MoE) structure that mixes three main improvements to ship as much as 5x increased throughput and as much as 2x increased accuracy than the earlier Nemotron Tremendous mannequin.
- Hybrid Structure: Mamba layers ship 4x increased reminiscence and compute effectivity, whereas transformer layers drive superior reasoning.
- MoE: Solely 12 billion of its 120 billion parameters are lively at inference.
- Latent MoE: A brand new approach that improves accuracy by activating 4 professional specialists for the price of one to generate the following token at inference.
- Multi-Token Prediction: Predicts a number of future phrases concurrently, leading to 3x sooner inference.
On the NVIDIA Blackwell platform, the mannequin runs in NVFP4 precision. That cuts reminiscence necessities and pushes inference as much as 4x sooner than FP8 on NVIDIA Hopper, with no loss in accuracy.
Open Weights, Information and Recipes
NVIDIA is releasing Nemotron 3 Tremendous with open weights underneath a permissive license. Builders can deploy and customise it on workstations, in knowledge facilities or within the cloud.
The mannequin was skilled on artificial knowledge generated utilizing frontier reasoning fashions. NVIDIA is publishing the whole methodology, together with over 10 trillion tokens of pre- and post-training datasets, 15 coaching environments for reinforcement studying and analysis recipes. Researchers can additional use the NVIDIA NeMo platform to fine-tune the mannequin or construct their very own.
Use in Agentic Techniques
Nemotron 3 Tremendous is designed to deal with complicated subtasks inside a multi-agent system.
A software program improvement agent can load a complete codebase into context directly, enabling end-to-end code technology and debugging with out doc segmentation.
In monetary evaluation it could actually load 1000’s of pages of experiences into reminiscence, eliminating the necessity to re-reason throughout lengthy conversations, which improves effectivity.
Nemotron 3 Tremendous has high-accuracy device calling that ensures autonomous brokers reliably navigate large perform libraries to stop execution errors in high-stakes environments, like autonomous safety orchestration in cybersecurity.
Availability
NVIDIA Nemotron 3 Tremendous, a part of the Nemotron 3 household, may be accessed at construct.nvidia.com, Perplexity, OpenRouter and Hugging Face. Dell Applied sciences is bringing the mannequin to the Dell Enterprise Hub on Hugging Face, optimized for on-premise deployment on the Dell AI Manufacturing unit, advancing multi-agent AI workflows. HPE can be bringing NVIDIA Nemotron to its brokers hub to assist guarantee scalable enterprise adoption of agentic AI.
Enterprises and builders can deploy the mannequin by a number of companions:
- Cloud Service Suppliers: Google Cloud’s Vertex AI and Oracle Cloud Infrastructure, and coming quickly to Amazon Internet Companies by Amazon Bedrock in addition to Microsoft Azure.
- NVIDIA Cloud Companions: Coreweave, Crusoe, Nebius and Collectively AI.
- Inference Service Suppliers: Baseten, Cloudflare, DeepInfra, Fireworks AI, Inference.web, Lightning AI, Modal and FriendliAI.
- Information Platforms and Companies: Distyl, Dataiku, DataRobot, Deloitte, EY and Tata Consultancy Companies.
The mannequin is packaged as an NVIDIA NIM microservice, permitting deployment from on-premises methods to the cloud.
Keep updated on agentic AI, NVIDIA Nemotron and extra by subscribing to NVIDIA AI information, becoming a member of the neighborhood, and following NVIDIA AI on LinkedIn, Instagram, X and Fb.
