NVIDIA: Nemotron 3 Ultra (free)

NVIDIA Nemotron 3 Ultra is the large-scale reasoning and orchestration model in the Nemotron 3 family. It is designed for demanding multi-step agents, long-context analysis and reasoning over code, mathematics and science. The model has 550B total and 55B active parameters. It combines Mamba-2, attention and Latent Mixture-of-Experts layers, with Multi-Token Prediction layers supporting its generation design.

LatentMoE projects information into a smaller latent representation for expert processing. NVIDIA’s training recipe combines pre-training, supervised fine-tuning, reinforcement learning and multi-domain on-policy distillation. The practical purpose of these stages is to improve reasoning, instruction following, tool use and agent behavior across several task types. Ultra produces text: an explanation, a solution, code, a plan or a requested response structure. Supply the material the answer should use and the conditions that determine whether the result is correct.

For mathematical or scientific reasoning, include the complete statement, known values, units and permitted assumptions. A useful request is: “Derive the result from these conditions, identify the assumptions at each stage, check the dimensions, and explain which observation would invalidate the conclusion.” Ask for a concise final result after the working explanation. If experimental data is involved, distinguish what was measured from what was calculated, and ask for uncertainty or missing information to remain visible rather than being absorbed into a confident answer.

For code and agent workflows, provide the objective, relevant files as text and the real outputs of prior steps. Ultra is intended for tasks involving several dependent decisions, so an explicit record of completed work and remaining questions is useful. For example: “Review this migration proposal against the schema and failing checks. Identify the dependency that blocks completion, propose a focused correction, and give the expected result of the next check.” A tool-connected application can then execute a proposed step and return evidence for the next decision.

The manufacturer supports reasoning on or off through the chat template. Reasoning is on by default, producing a reasoning trace followed by the answer. The Ultra card also documents a medium-effort reasoning mode that uses fewer reasoning tokens than full thinking. This is an Ultra-specific option; it should not be confused with the low-effort setting documented for Super. Set the desired final deliverable and leave space for it after the reasoning. An unfinished trace does not substitute for a completed answer to the original question.

Ultra supports up to a 1M-token context. For a large collection of documents or source files, label sections and tell the model exactly what to compare, trace or reconcile. Keep the source order when chronology or dependency matters. A useful long-document result contains the finding, the supporting passage and the implication for the question. Select the relevant materials and reserve output space; the context window describes the working sequence rather than a separate guaranteed final-answer length.

The maker lists English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Brazilian Portuguese and Chinese. State the output language and preserve technical identifiers or source quotations where exact wording matters. For a multilingual analysis, use the same criteria for each document and ask for differences in terminology or assumptions to be explained. Fluency alone is not enough to establish that two differently worded requirements mean the same thing.

NVIDIA provides explicit tool-calling examples with reasoning. Define each function’s purpose, arguments and expected result, then return its actual output to the model. The application performs the operation; the model reasons over the returned observation. A long agent workflow becomes easier to inspect when proposed actions, executed results and unresolved checks are clearly distinguished. For a final report, request the answer, the evidence that supports it, the assumptions and any remaining task items. Verify calculations against the original units, code changes against observed checks, and factual claims against the supplied sources.