NVIDIA Nemotron 3
NVIDIA's open-weight family, 31.6B Nano up to the 550B Ultra MoE. Hybrid Mamba-Transformer, 1M context, weights plus roughly 3T tokens of training data published.
Overview
Nemotron 3 is NVIDIA's open-weight line, structured as three sizes plus a speed variant. Nano at 31.6B total and 3.6B active arrived 15 December 2025. Super, around 120B with 12.7B active, followed at GTC in March 2026. Ultra at 550B total and 55B active landed 4 June 2026, pretrained on roughly 20T tokens with a 1M token context. All of them use a hybrid Mamba-Transformer design with latent mixture-of-experts routing, and Super and Ultra were trained in NVIDIA's 4-bit NVFP4 format on Blackwell.
Nemotron 3.5 Lightning, out 11 August 2026, is the one most engineers will actually run. Same 31.6B total and 3.6B active shape as Nano but rebuilt for the execution layer of long-running agents: an intelligence index of 24, output speeds near 670 tokens per second on DeepInfra endpoints and a claim of finishing 10,000 PinchBench tasks 30% faster than a comparable 35B open model at similar accuracy. NVIDIA shipped NeMo Switchyard with it, a router that sends simple work to Lightning and escalates the rest.
The licence question deserves a straight answer. The Nemotron 3 models were released under the NVIDIA Open Model License, a permissive vendor licence allowing commercial use, modification and redistribution with attribution and standard patent-termination and trade-compliance terms, but not an OSI-approved open source licence. Newer releases including Ultra and Lightning use OpenMDW-1.1, the Linux Foundation licence published 28 May 2026 that covers weights, data, code and evaluation artefacts under one agreement. The Linux Foundation submitted OpenMDW to the OSI for certification in mid-August 2026. It has not been approved. Treat these as open weights with an unusually generous data release, not as open source software.
Key Features
- ✓ Four current members: Nemotron 3 Nano at 31.6B total and 3.6B active from 15 December 2025, Super at roughly 120B with 12.7B active from March 2026, Ultra at 550B with 55B active from 4 June 2026 and Nemotron 3.5 Lightning at 31.6B A3.6B from 11 August 2026
- ✓ Hybrid Mamba-Transformer with latent mixture-of-experts routing, 1M token context on Ultra and Lightning
- ✓ BF16 and NVFP4 checkpoints published; Super and Ultra were trained in 4-bit NVFP4 on Blackwell
- ✓ Roughly three trillion tokens of pretraining, post-training and reinforcement learning data released alongside the weights, plus an agentic safety dataset
- ✓ NeMo Gym, NeMo RL and NeMo Evaluator published as open training and evaluation infrastructure
- ✓ NeMo Switchyard routes traffic between Lightning and the heavier models so cheap requests do not hit the expensive one
Where it holds
- • The data release is not decoration. Three trillion tokens of pretraining and RL data means post-training is reproducible rather than guesswork, which almost no other vendor family offers
- • Nemotron 3.5 Lightning scores 24 on the Artificial Analysis intelligence index at 31.6B total, roughly matching gpt-oss-120b at about a quarter the parameters, and serves near 670 tokens per second on DeepInfra
- • OpenMDW-1.1 covers weights, data and code in one agreement instead of the usual patchwork, and the Linux Foundation submitted it to the OSI for certification in mid-August 2026
- • Deployment targets are broad within the NVIDIA stack: Blackwell, Hopper and Ampere, scaling from Jetson and DGX Spark up to data centre, with an RTX 5090 handling Lightning locally
Where it breaks
- • Licensing is split across the family. The original Nemotron 3 models shipped under the NVIDIA Open Model License, newer releases moved to OpenMDW-1.1, so read the licence file on each repo rather than assuming
- • OpenMDW-1.1 is not OSI-approved. It went in for review around 14 August 2026 and is still pending, so calling it open source is premature
- • Everything is tuned for NVIDIA silicon. NVFP4 needs Blackwell, and the throughput claims that make Lightning interesting do not survive a move to other accelerators
- • The naming is a mess. Nano, Super, Ultra and Lightning across two version numbers in flight at once, and the marketing benchmarks are all framed as speed per token because the point is selling GPUs
My Take
Hardware vendors publish models to sell accelerators, and the benchmark framing here is almost entirely speed per token, so read the marketing with that in mind. The engineering underneath holds up anyway: Nemotron 3.5 Lightning, released 11 August 2026 at 31.6B total and 3.6B active, hits an Artificial Analysis index of 24 and quantises onto one RTX 5090, and NVIDIA shipped around 3T tokens of training data with it, which puts it well ahead of Gemma or Muse on transparency. The licence needs care. Nemotron 3 came out under the NVIDIA Open Model License, Ultra and Lightning use OpenMDW-1.1, and OSI review of OpenMDW only started in mid-August 2026, so it is permissive but not yet certified open source.
Quick Info
- Pricing:
- open-source
- Openness:
- Open weights
- Licence:
- OpenMDW-1.1
- Starting at:
- Weights are free from Hugging Face, ModelScope and build.nvidia.com, in BF16 and NVFP4. Hosted, Nemotron 3.5 Lightning is $0.05 in and $0.20 out per million tokens on DeepInfra and $0.10 in, $0.25 out on CoreWeave, both routed through OpenRouter, which also runs a free tier capped at 65,536 completion tokens. Checked 28 August 2026. NVIDIA sells the supported NIM runtime separately through NVIDIA AI Enterprise, which is not required to run the weights yourself.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where NVIDIA Nemotron 3 holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to NVIDIA Nemotron 3
Google Gemma 4
open-sourceGoogle's open-weight family in five sizes, 2B effective up to 31B dense, text and image in with audio on the smaller ones, 256K context and Apache 2.0 weights since April 2026.
Mistral AI
open-sourceEuropean lab publishing open weight models you can run on your own hardware. Large 3 is a 675B MoE with 41B active, 256K context and Apache 2.0 weights.
Ai2 Olmo 3
open-sourceAi2's fully open family at 7B and 32B with Base, Instruct, Think and RL Zero variants. Weights, training data, every intermediate checkpoint, logs and code under Apache 2.0.
MiniMax M3
freemiumMiniMax's open-weight flagship LLM (稀宇科技), distinct from the Hailuo video product: 428B mixture-of-experts, 23B active per token, 1M context and native image and video input.