NVIDIA Nemotron 3 in Tarsk

NVIDIA shipped Nemotron 3 in three sizes: Nano (30B), Super (120B), and Ultra (550B). All three are open-weight MoE models trained for tool use and long context. Tarsk already lists Nemotron 3 endpoints from nine providers, so you can point an agent at one without writing a custom integration.

The model line

Nemotron 3 targets agent workloads: multi-step tool calls, long files in context, retries when a command fails. NVIDIA released weights, training data (where redistribution is allowed), and recipes under the Open Model License. You can self-host or call a hosted endpoint.

The backbone mixes Mamba state-space layers with Transformer attention in a Mixture-of-Experts layout. Each forward pass activates a small slice of total parameters. Nano runs about 3B of 30B. Ultra runs about 55B of 550B.

Nano, Super, Ultra

• Nano (30B total, ~3B active) fits tight inference budgets. NVIDIA reports 4× higher throughput than Nemotron 2 Nano on comparable hardware. Good fit for debug loops, summarization, and retrieval-heavy agents.
• Super (120B total, ~12B active) adds LatentMoE routing, Multi-Token Prediction, and NVFP4 training. NVIDIA positions it for multi-agent setups like ticket routing and IT automation.
• Ultra (550B total, ~55B active) landed at Computex 2026. Same MoE tricks at frontier scale. NVIDIA cites up to 5× throughput versus the prior generation and roughly 30% lower operating cost on Blackwell hardware.

What changes for your agents

Chat billing counts tokens per turn. Agent billing counts tokens across every tool call, planning step, and retry. Nemotron 3 optimizes for that second pattern.

• MoE routing limits active parameters per token so a 120B model does not behave like a dense 120B model at inference time.
• Multi-Token Prediction generates several tokens per step, which speeds long outputs and supports speculative decoding without a separate draft model.
• Mamba layers scale context length closer to linear cost than standard attention, which matters when you paste a whole repo into the prompt.
• NeMo Gym and NeMo RL give you post-training environments if you need a domain-specific agent on top of the base checkpoints.

Providers in Tarsk

Tarsk syncs provider catalogs on a schedule. When OpenRouter or Together AI lists a new Nemotron endpoint, it shows up under Settings → Models without waiting for a Tarsk release.

First-party and local

• NVIDIA at build.nvidia.com: Nano, Super, Ultra, plus content-safety and embedding variants.
• Ollama Cloud: nemotron-3-nano:30b and nemotron-3-super.

Ultra (recent catalog additions)

• OpenRouter: nvidia/nemotron-3-ultra-550b-a55b, a free tier, and a content-safety model.
• Together AI: nvidia/nemotron-3-ultra-550b-a55b.
• OpenCode Zen: nemotron-3-ultra-free (replaced the deprecated Super free tier).

Nano and Super

• Kilo Gateway: Nano 30B, Super 120B (paid and free), Nano Omni reasoning.
• NanoGPT: Super with a thinking variant, Nano, Nano Omni reasoning.
• Nebius Token Factory: Nano Omni, Super 120B, Nano 30B.
• Cortecs: nemotron-3-super-120b-a12b.

Try it in Tarsk

1. Add a provider key under Settings → Providers (NVIDIA, OpenRouter, or Together AI are the fastest paths to Ultra).
2. Open Settings → Models, search nemotron, enable the tier you want.
3. Pick that model in your thread and send a task with tools turned on.

Start with Nano if you care about cost per step. Move to Super for production agent fleets. Use Ultra when a task needs long reasoning chains and you have the budget for 550B-class inference.

Summary