Uncensored Nemotron API

Most "open" models are open weights and nothing else — a tarball of parameters, a license, and silence about where the numbers came from. Nemotron has been the consistent exception for two years: NVIDIA ships the weights, the post-training datasets, the RL environments and the recipe. That matters for an uncensored endpoint in a way that is easy to miss. A model whose alignment data is published is one whose refusal behavior you can reason about instead of guess at — and one a third party can legitimately deploy without a content filter bolted to its mouth.

01What is Nemotron 3 Ultra?

Nemotron 3 Ultra is the largest member of NVIDIA's Nemotron 3 family, released to Hugging Face on June 4, 2026. NVIDIA's technical report describes it as "a 550 billion total and 55 billion active parameter Mixture-of-Experts Hybrid Mamba-Attention language model" — roughly 10% of the network works on any given token, which is what makes a half-trillion-parameter model tractable to serve at all.

The reported layout is 108 layers at model dimension 8,192, interleaving Mamba-2 state-space layers with a sparse set of full attention layers, 512 experts per MoE layer with top-22 routing. Pretraining ran on roughly 20 trillion text tokens in two phases — diversity-biased, then quality-biased — with a September 2025 cutoff, 12 natural languages and 43 programming languages. Post-training used SFT, multi-environment RL with verifiable rewards, and Multi-Teacher On-Policy Distillation across more than ten specialist teachers. The stated design target is long-running agents: sessions that hold a large context, call tools dozens of times, and run long enough that per-token economics dominate the bill.

02The Nemotron lineage, and why the pedigree matters

Nemotron is not a one-off. It is the visible output of an NVIDIA research program running in the same direction since 2024, and each generation left something the current model still carries. Nemotron-4 340B (June 2024) shipped with an open synthetic data generation pipeline, its reward model and its generator, plus an unusual disclosure: the overwhelming majority of its alignment data was model-generated. The model as a data factory rather than an oracle is why later Nemotron releases keep shipping datasets alongside weights.

Llama-Nemotron — reasoning you can switch off

The 2025 Llama-Nemotron family (LN-Nano 8B, LN-Super 49B, LN-Ultra 253B) contributed two ideas that aged well. The first was architectural: NVIDIA's Puzzle neural architecture search built a library of block-wise-distilled transformer blocks and assembled them with mixed-integer programming, producing models where some blocks skip attention entirely or replace it with a single linear layer, while FFN Fusion collapsed consecutive feed-forward blocks into fewer, wider, parallel ones — cutting sequential depth rather than parameter count. LN-Ultra fit on a single 8×H100 node while, per NVIDIA's paper, outperforming DeepSeek-R1 on GPQA-Diamond (76.0% vs 71.5%).

The second was behavioral: Llama-Nemotron was the first open family with a dynamic reasoning toggle, tagging training data detailed thinking on or detailed thinking off in the system instruction, with format rewards enforcing <think> tags in the on-mode. One set of weights, two behaviors, selectable at inference time. Nemotron 3's reasoning budget control — and the reasoning_effort parameter on our API — descends directly from that work.

Nemotron 3 — the hybrid turn

Nemotron 3 replaced the Llama derivation with a from-scratch hybrid architecture and added LatentMoE (compressing tokens before routing, which NVIDIA says allows "4x as many experts for the same inference cost"), multi-token prediction, and native NVFP4 pretraining. NVIDIA published a cumulative 50 million SFT samples, 2 million RL tasks and 55 RL environments alongside the weights — you do not have to take a vendor's word for what the model was taught, because the teaching material is downloadable.

03What the hybrid architecture buys you

A transformer's KV cache grows linearly with sequence length, which is why long-context inference gets expensive in memory before compute. Mamba-2 layers are state-space models carrying a fixed-size recurrent state, so their per-token memory cost does not grow with history; what they are bad at is precise associative recall. The hybrid keeps Mamba-2 for the bulk of sequence processing plus a minority of real attention layers for the recall state-space layers fumble, with an aggressive 64-query-head / 2-KV-head ratio shrinking the cache further. NVIDIA reports 94.7% on RULER at 1M tokens — a needle-retrieval benchmark built to catch models claiming a window they cannot use.

The consequence is throughput: NVIDIA claims up to roughly 6× higher inference throughput than comparable models at on-par accuracy, and Artificial Analysis independently measured 140.3 tokens/second with a 1.33-second time to first token — 7th of 89 models. On a 200-turn agent run that is the difference between finishing in an afternoon and not finishing.

04What "3BIT-REAP" means, honestly

The suffix is not decoration — it tells you exactly what was done to the foundation to make it fit.

REAP — Router-weighted Expert Activation Pruning — comes from Cerebras Research and was accepted to ICLR 2026. In a sparse MoE the router sends each token to a handful of experts; over a real workload some are hit constantly and others almost never. REAP scores each expert on a saliency criterion combining how strongly the router gates it and how large its output contributions are, then removes the least salient outright — Cerebras reports pruning up to 50% of experts from models as large as 1 trillion parameters while largely preserving baseline quality, and finds that for generative tasks like code, pruning beats merging. 3-bit quantization then compresses what survives. Prune first, quantize second, so the bit budget goes to experts that carry traffic instead of being smeared across dead ones.

Compression is not free, and we will not pretend otherwise: aggressive quantization shows up first as small degradations in exact-format compliance and long-chain arithmetic, not as visible incoherence. Where the three tiers land:

Suffix	What it is	Where we use it
3BIT-REAP	3-bit weights + expert pruning	The largest foundations, where nothing else fits at speed
W4A16-AUTOROUND-REAP	4-bit weights / 16-bit activations, AutoRound learned rounding, + pruning	Mid-size foundations; AutoRound recovers accuracy over naive round-to-nearest
W8A16	8-bit weights / 16-bit activations	Highest fidelity; foundations small enough not to need aggressive compression

Nemotron 3 Ultra at 550B is squarely in the first category. If your workload is unusually sensitive to exact numeric output over long chains, the research index lists the W8A16 tier — MiMo-V2.5-Pro-W8A16 and Hy3-W8A16 — at the $0.50/$2.00 floor. A real trade, not a funnel: smaller model, less compression.

05What "uncensored" means on this endpoint

Concretely: there is no refusal layer and no content filtering applied to output. No classifier inspects your prompt and returns a canned decline, and the response is not scanned before it reaches your stream. What you get is the model's own behavior.

This is where the open lineage stops being trivia. Because NVIDIA published the post-training data and the RL environments, the model's dispositions are auditable in principle rather than folklore. And because OpenMDW-1.1 is a permissive Linux Foundation license covering architecture, parameters, documentation and software, an independent operator can serve a modified, unfiltered deployment on their own hardware without the license being the obstacle. Many "open" model licenses do not permit that. This one does. The difference shows up where reflexive refusal is a bug, not a safety feature:

• Security research — vulnerability analysis, exploit reasoning, malware behavior, red-team tooling, without arguing about whether the question is allowed.
• Chemistry, biology and pharmacology — dosage math, reaction mechanisms, toxicology, synthesis routes that a filtered endpoint declines on keyword contact.
• Adversarial ML — hard negatives, jailbreak corpora and abuse taxonomies for classifiers that need to have seen the real thing.
• Clinical, legal and forensic text — case material that is graphic by nature and gets refused for exactly that reason.

Uncensored is not unaccountable. You own what you generate, and our Responsible Use Policy and Terms of Service govern the account. Removing a refusal layer removes a guess about intent; it does not remove your obligations under the law that applies to you.

06STEM and reasoning: what the published numbers say

Nemotron's training focus has been math, code, science and instruction following since the Llama-Nemotron post-training dataset was built around exactly those four axes. What NVIDIA reports for the Nemotron 3 Ultra foundation — vendor-reported figures on the uncompressed model, which we did not run:

Two caveats worth more than the numbers. The SWE-Bench Verified figure is a range — NVIDIA reports 65.0% to 70.4% depending on which agent harness runs it (Pi, OpenHands, Hermes, OpenCode, Mini SWE Agent), and a five-point spread from scaffolding alone shows how much agentic scores depend on the harness rather than the model. And Artificial Analysis placed Nemotron 3 Ultra at 48 on its Intelligence Index, 9th of 89 models — strong, not top of the field, and their independent measurement is the one we would weight over any vendor's slide.

On the earlier generation, NVIDIA's Llama-Nemotron paper reports LN-Ultra at 76.0% on GPQA-Diamond against DeepSeek-R1's 71.5% and Llama-3.1-405B-Instruct's 43.4%, plus 80.8% on AIME 2024 — a different model, included because graduate-level science and competition math have been the family's target for two generations.

Benchmark figures as published by NVIDIA (June 2026) on the uncompressed Nemotron 3 Ultra, and by Artificial Analysis for independent measurements. Compressed builds may differ. We do not publish benchmark numbers we did not verify or cannot attribute.

07Price, and the verbosity caveat

At $0.75 per million input tokens and $3.30 per million output tokens, Nemotron3Ultra sits in the lower third of our catalog on input and the middle on output. Worked examples, because per-million pricing is hard to feel:

Call shape	Input	Output	Cost
Typical RAG turn	50,000	4,000	$0.051
Long document analysis	150,000	8,000	$0.139
Maximum-size call	262,144	16,384	$0.251

A full-context, full-output request — every token the endpoint will accept and every token it will produce — costs about 25 cents. The same shape on Kimi-K3-3BIT-REAP, our most expensive model at $3.83/$19.12, costs about $1.32, or 5.3× more.

Now the caveat, and it is real. Artificial Analysis measured Nemotron 3 Ultra generating 2.3× more output tokens than the median peer across their benchmark suite. Reasoning models bill you for the thinking, and this one thinks at length; a per-token price advantage evaporates into a per-task disadvantage if the model reasons three times as long to reach the same answer. The mitigation is built in and exposed on our API: Nemotron 3 was post-trained for reasoning budget control, and reasoning_effort is supported here — NVIDIA describes a lower-effort mode using substantially fewer thinking tokens for a single-digit accuracy trade-off. On classification, extraction and routing, turn it down and the economics change materially; on a proof or a multi-file refactor, turn it up and pay for it deliberately. Measure your own tokens-per-task rather than reading the price list and assuming.

08Calling it: three dialects, one model

Every model on the gateway answers on three streaming API surfaces, so you can point an existing client at it without rewriting your integration: /v1/chat/completions (OpenAI dialect), /v1/messages (Anthropic dialect), and /v1/responses.

curl https://api.shannon-ai.com/v1/chat/completions \
  -H "Authorization: Bearer $SHANNON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Nemotron3Ultra-3BIT-REAP",
    "reasoning_effort": "medium",
    "stream": true,
    "messages": [
      {"role": "user", "content": "Derive the worst-case cache behavior of this loop nest."}
    ]
  }'

Capability summary, so nothing surprises you at integration time:

Capability	Nemotron3Ultra · 3BIT-REAP
Context window	262,144 tokens
Max output / request	16,384 tokens
Streaming	Yes
Tool calling	Yes
Reasoning traces	Yes
reasoning_effort	Yes
json_schema (strict)	Yes
response_format	Yes
Vision / image input	No — text only
Refusal layer	None

The vision row catches people: Nemotron3Ultra is text only. If your pipeline sends images, use MiniMax-M3-3BIT-REAP or MiMo-V2.5-W8A16 at $0.50/$2.00, or Kimi-K2.6-W4A16-AUTOROUND-REAP at $0.78/$3.67 — all vision-capable, all uncensored, all on the same 262,144-token context. Full parameter reference is in the API documentation.

09Where Nemotron3Ultra fits among the twelve

The catalog is twelve uncensored models sharing the same context window, the same three dialects and the same absence of a refusal layer; they differ in foundation, price and capability. Nemotron's position is specific: the open-lineage reasoning workhorse — not the cheapest, not the most capable, but the one with the most public paper trail behind its behavior, tuned hardest toward STEM and agents.

Model	In / 1M	Out / 1M	Vision	Pick it for
MiniMax-M3 · 3BIT-REAP	$0.50	$2.00	Yes	Cheapest vision-capable option
MiMo-V2.5-Pro · W8A16	$0.50	$2.00	No	Highest-fidelity quantization tier
GLM-5.2 · 3BIT-REAP	$0.73	$2.34	No	Cheapest large-foundation reasoning
Nemotron3Ultra · 3BIT-REAP	$0.75	$3.30	No	STEM, agents, open lineage
Kimi-K2.6 · W4A16-AUTOROUND-REAP	$0.78	$3.67	Yes	Vision + AutoRound fidelity
DeepSeek-V4-Pro · 3BIT-REAP	$1.95	$3.90	No	Deep single-shot analysis
Kimi-K3 · 3BIT-REAP	$3.83	$19.12	Yes	Absolute capability ceiling

Seven of twelve shown. The full catalog — including Laguna-S-2.1, inkling, DeepSeek-V4-Flash, MiMo-V2.5 and Hy3 — is in the API docs. Note that Laguna-S-2.1 does not support structured output and inkling does not support json_schema; Nemotron3Ultra supports both.

The closest neighbor by price is GLM-5.2-3BIT-REAP at $0.73/$2.34, and the honest comparison is this: GLM-5.2 is cheaper per token, especially on output. Nemotron brings a published training corpus, an explicit reasoning-budget dial, and a family history aimed at graduate-level science. High-volume general reasoning: run the cheaper one. STEM-heavy, agentic, or subject to an audit that asks where the model's behavior came from: Nemotron earns the difference.

10Frequently asked questions

---

Sources: NVIDIA Nemotron 3 Ultra technical report · NVIDIA developer blog (June 2026) · Nemotron 3 Super architecture notes · Llama-Nemotron: Efficient Reasoning Models · Cerebras Research, REAP. Foundation benchmarks are NVIDIA-reported on the uncompressed model unless attributed to Artificial Analysis; compressed builds may differ. Prices and capabilities are current as of September 5, 2026.