NVIDIA Nemotron 3 Free Guide: Call $0 on OpenRouter, 550B Flagship + 30B Multimodal, up to 1M Context

NVIDIA's Nemotron 3 model family has just landed in OpenRouter's :free pool, and this is not a toy model, it is a flagship 550B-parameter one. The key numbers (verified 2026-10-04 against the live OpenRouter API): $0 input / $0 output, up to 1,000,000-token context, and these :free variants currently have no expiration date (expiration is empty), unlike Space Bunny, which went down on 10/05. The family spans four tiers: Ultra 550B (55B active) is the flagship reasoning/agent tier, Super 120B (12B active) is the balanced tier, Nano Omni 30B (~3B active) is the only multimodal tier (ingests video/audio/image/text directly), and 3.5 Lightning is a fast 1M-context tier. Bottom line up front: if you want a $0 way to run NVIDIA's newest Mamba-2 + MoE hybrid architecture with million-scale long context, OpenRouter's Nemotron 3 free pool is the most concrete entry right now, but mind OpenRouter's hard daily request cap for free tiers (a never-topped-up account gets only 50 requests/day); do not mistake "free" for "unlimited". Below are all the measured numbers, usage, and the gotchas.

Free Allowance & Specs at a Glance

All five variants are $0 in / $0 out and currently have no expiration. Other free models sharing the same free pool you can mix across: google/gemma-4-26b-a4b-it:free and google/gemma-4-31b-it:free (262K context, image/text/video input), qwen/qwen3.8-27b:free (image/text/video input), cohere/north-mini-code:free (256K, coding-focused), and dots-studio/dots-3-note-preview:free (512K, image in, expires 2026-12-31).

OpenRouter Free Rate Limits: The Real Barrier

"$0" does not mean "unlimited". OpenRouter gates :free variants by your all-time top-up credits in two tiers (2026-10-04 docs):

• The ceiling applies to free-model requests, not paid ones; the 20 req/min cap applies to all free requests;
• To see your live allowance, call GET /api/v1/key and read the free_model_daily_requests counter and ceiling;
• A never-topped-up account gets 50/day; heavy Ultra 550B long-context use will hit that fast, and extra accounts do not help, OpenRouter governs capacity globally per account/key;
• On a 429, back off exponentially and honor the Retry-After header.

How to Use It at $0 (From Key to Call)

1. Register at openrouter.ai and create an API key (free, no card required);
2. Write the model id with the :free suffix, e.g. nvidia/nemotron-3-ultra-550b-a55b:free;
3. Minimal call (OpenAI-compatible, Python): curl https://openrouter.ai/api/v1/chat/completions -H "Authorization: Bearer YOUR_KEY" -d '{"model":"nvidia/nemotron-3-ultra-550b-a55b:free","messages":[{"role":"user","content":"..."}]}'
4. For long context: drop an entire codebase / paper / log into the 1M-token window and let Ultra do cross-file analysis;
5. For multimodal: send video + audio + image + text to Nano Omni 30B and get transcription / document intelligence / GUI analysis back;
6. When you are tight on quota, switch the same job across nemotron-3-super-120b, gemma-4-31b, and nemotron-3.5-lightning as fallbacks to preserve the flagship tier's 50/day.

Nemotron 3 Architecture: Why 550B Is Given Away

• LatentMoE: interleaved Mamba-2 and MoE layers plus selective Attention, with Multi-Token Prediction (MTP) heads doing shared-weight speculative decoding for faster inference;
• NVFP4 pre-training recipe: weights/activations/gradients mostly in NVFP4 quantization-aware training to cut compute;
• Ultra 550B was pre-trained on roughly 20T tokens and refined via Multi-Domain On-Policy Distillation (MOPD), distilling strong teachers onto its own rollouts to lift code / math / tool-use / agentic behavior;
• Nano Omni 30B is a Mamba2-Transformer hybrid MoE that unifies video/audio/image/text understanding for enterprise Q&A, transcription, OCR, document intelligence, and GUI agents.

Five $0 Use Cases (Pick by Scenario)

1. Million-token document / codebase analysis: Ultra 550B's 1M window + switchable reasoning for cross-file RAG and whole-repo review;
2. Multimodal content understanding: Nano Omni 30B eats video/audio/image/text for meeting transcription, dense captions, GUI analysis;
3. Fast long-context iteration: 3.5 Lightning shares the 1M context but is lighter, good for multi-turn agents;
4. Balanced value: Super 120B (12B active) at 262K for everyday Q&A and instruction following, uses less quota;
5. Safety pre-check: the content-safety variant (128K, text+image) for input filtering and risk assessment.

Free-Entry Comparison

(All figures from the OpenRouter models API, its docs, and NVIDIA HuggingFace model cards; allowances and expiration dates follow OpenRouter's live in-app display.)

FAQ

Bottom Line

Nemotron 3's free policy is refreshingly numeric: five OpenRouter tiers all $0 (Ultra 550B / Super 120B / Nano Omni 30B multimodal / 3.5 Lightning / content-safety), up to 1M context, no expiration today; the real constraint is the free request ceiling (under $10 top-up: 50/day; $10+: 1000/day; 20/min). If you want a $0 1M-context or $0 multimodal run, this is the most stable entry; for unlimited self-hosting, use NVIDIA's open weights. Allowances and expiration dates are in flux, verify in OpenRouter's app before committing.