Beyond GPUs, NVIDIA ships the Nemotron family of open-weight models — designed to be post-trained on your own data for enterprise agents, legal, search, clinical and sovereign-AI workloads at a fraction of closed-frontier cost.
NVIDIA’s Nemotron Labs pitches open models as the customization layer for enterprises: keep full control of weights, inspect training, run private evals and stand up agents on your own infrastructure. Deployments include Harvey (legal), Glean (agentic search, Waldo), H Company (computer use) and YTL AI Labs (Malay sovereign AI).
| Model | What it is for | Vendor list price |
|---|---|---|
| Nemotron 3 Ultraopen flagship | Frontier-class open model for enterprise post-training; top open-model agent accuracy via LangChain Deep Agents; Harvey matched legal frontier at ~10x lower cost | Open weights — free to self-host |
| Nemotron 3 Nano Omnismall open multimodal | Small-model computer-use: H Company’s Holotron 3 Nano scored over 76% on OSWorld-Verified | Open weights — free to self-host |
| Nemotron 3.5 Lightning30B MoE / 3B active · 11 Aug 2026 | Resident-agent workhorse — ~4x faster inference, paired with the open NeMo Switchyard routing library | Open weights — free to self-host |
Last updated — Nemotron is open-weight: the model licence is free for self-hosting. Because there is no vendor list price, the cost that matters is whatever your host charges — a post-trained deployment’s per-token cost depends on the hardware and the batch size, and the third-party figures in circulation vary too widely to quote as a single rate. Hosted enterprise access runs via NVIDIA NIM / build.nvidia; self-hosted cost is your own GPU time. Figures are vendor public list prices in USD per 1M tokens, standard tier — no batch, cache, promo or peak-hour discount applied. Vendors change list prices without notice: confirm on the vendor’s own pricing page before budgeting.
Discontinued or fully superseded models, kept for historical reference. Active models — even previous-generation ones still on sale — always stay in the lineup above.
| Model | Era & what it was for | Historic list price |
|---|---|---|
| Llama-3.3-Nemotron-Super-49BMar 2025 · open distillation | The 2025 open enterprise workhorse distilled from Llama 3.3 70B — reasoning and tool-use at a single-GPU-friendly size | Open weights — free to self-host |
| Llama-3.1-Nemotron-70B2024 · NVIDIA-refined Llama | NVIDIA’s first widely adopted refined-Llama open model — strong instruction following for its era | Open weights — free to self-host |
Recent announcements with outlet links (EN first, CN where available).
Nemotron Labs listed six production deployments — Harvey, Glean, H Company, Abridge, Heidi Health and YTL AI Labs — matching closed-frontier accuracy at 10–20x lower cost per run.
EN · AI Chat Daily ↗A 30B MoE (3B active) resident-agent model at ~4x the speed, plus an open-source routing library for multi-model agent setups.
EN · NVIDIA developer blog ↗CN · NVIDIA blog (CN) ↗Jointly built on Nemotron open models and trained on Malaysian data on YTL AI Cloud — a flagship example of the open-model sovereign-AI pattern.
EN · Bernama ↗CN · Media Selangor ↗One partner for every digital layer – explore what else we build, host and grow for Malaysian businesses.
