Model developer guide
5 website model guides
Choose Nemotron 3.5 Super VL for visual analysis or compare Ultra, Super, Lightning and Nano for text-focused reasoning. Each has a distinct paid OpenRouter route. NVIDIA’s underlying model specifications, the serving host’s limits and the app’s enabled controls all matter when choosing a workflow.
Practical guidance
Editorial starting points, not a performance ranking. Check current app access before choosing a model.
Use the vision-language guide for a written question tied to readable visual evidence. NVIDIA documents text, image and video inputs, visual-document understanding and a 120B-total, 12B-active reasoning architecture.
Tradeoff: The underlying 1M-token capability is not a universal hosted allowance. Check the selected serving endpoint’s context and completion limits, and verify extracted values against the source.
Nemotron 3.5 Super VL guideEvaluate Ultra for a substantial text or code brief with competing constraints. Its NVIDIA card documents a text model, while OpenRouter exposes the named paid route and model-specific request capabilities.
Tradeoff: A large model's positioning is not a measured quality win on your workload. Compare the evidence quality and review effort, not only model size or a benchmark headline.
Nemotron 3 Ultra guideUse the Nano guide for a smaller repeatable task such as extracting facts from a short record. OpenRouter documents its separate nvidia/nemotron-3-nano-30b-a3b route and request fields.
Tradeoff: A compact starting point still needs difficult test cases and an escalation rule. Neither a successful small example nor a family label proves suitability for ambiguous long-form work.
Nemotron 3 Nano guideTechnical details
Official documentation describes the named model or gateway—not every feature available in this app.
The model card describes a 120B-total, 12B-active vision-language model with context up to 1M tokens. The reviewed OpenRouter endpoints include 262,144-token hosts and a 1,048,576-token host with different prices and completion ceilings. Reasoning and other request parameters depend on the selected hosted endpoint.
NVIDIA's Ultra card describes text input and maximum context up to 1M tokens in suitable deployments. The existing website's owner-confirmed OpenRouter coverage uses a 256K route reference, with host-specific differences documented in each model guide. Do not replace a served endpoint limit with the larger underlying model specification.
OpenRouter documents nvidia/nemotron-3-ultra-550b-a55b and nvidia/nemotron-3-nano-30b-a3b as separate model IDs on its chat-completions interface. The named routes do not include :free. A free-suffixed variant or another checkpoint would be a different selection, not an interchangeable spelling.
OpenRouter lists reasoning, tool and response-format controls for these model routes. NVIDIA also documents self-hosted chat-template requirements for reasoning and tool parsing. A self-hosting flag is not automatically an OpenRouter parameter, and advertised tool support does not execute an action without an application tool loop.
Complete website coverage
Every existing catalog entry for this provider is linked below. Status labels come from the individual website guides; catalog coverage is not a guarantee of current app access. Historical and snapshot pages remain references, not extra active models.
Practical guidance
Original prompt examples, not generated-output demonstrations. They have not been generation-tested.
Workflow 1
Provide two short proposals, the operational constraints and known failure history. Ask for a reviewable comparison before giving any agent permission to modify files or run commands.
Compare these database migration proposals using the supplied uptime requirement, rollback constraints and incident history. For every claimed risk, cite the supporting passage and describe a concrete failure scenario. Separate reversible experiments from irreversible steps. Recommend the smallest test that could change our choice. Do not invent measurements, tool output or approval from stakeholders.
Check the result: Verify the evidence and challenge unsupported risks. A proposed rollback is useful only if the actual system can perform and test it.
Read the Ultra guideWorkflow 2
Start with a short non-sensitive message and a written field schema. A prompt requesting JSON is not the same as an enforced schema; validate the response separately.
From the maintenance note below, return one JSON object with asset, observed_problem, reported_date and requested_action. Use null for an absent value. Preserve the author's wording for the problem and action. Include a separate evidence object quoting the exact supporting phrase for each non-null field. Do not infer dates from today's date or invent missing asset identifiers.
Check the result: Parse the output, check every quote and test messages with missing, contradictory or repeated fields. Keep failures visible instead of silently filling them in.
Read the Nano guideTechnical details
NVIDIA's official cards and OpenRouter endpoint documentation support the model identities, visual capabilities and hosting limits below. The four earlier mappings retain their existing owner confirmation. Host-specific reference limits and prices are not universal allowances or measured generation results.
Website catalog and OpenRouter model references
Super VL uses nvidia/nemotron-3.5-super-vl. The website retains Ultra as nvidia/nemotron-3-ultra-550b-a55b, Super as nvidia/nemotron-3-super-120b-a12b, Lightning as nvidia/nemotron-3.5-lightning and Nano as nvidia/nemotron-3-nano-30b-a3b. These are separate model IDs, not :free aliases or interchangeable spellings.
Existing model guides and NVIDIA model cards
The four earlier guides retain their text input/output and 262,144-token OpenRouter references. Super VL adds visual understanding through its own model identity. Its hosted endpoints can have different context ceilings, so reserve space for visual tokens, conversation history, reasoning and output.
OpenRouter endpoint documentation
Use the selected endpoint’s supported reasoning, tool and response-format parameters. NVIDIA’s chat-template flags describe its deployment interface rather than a universal gateway request. A named hosting price in a guide is a developer reference; customer plan allowances and usage charges are separate.
Before you rely on a result
Provider references and picker estimates are not EZ Ai Assist customer charges. Read the website pricing explanation.
Compare by task
Compare hosted and open-weight text-model options, keeping the checkpoint and serving contract explicit.
Evaluate another reasoning family against the same architecture and extraction tests.
Compare a serving platform's available text workflows while checking access and lifecycle conditions.
Common questions
The existing owner confirmation identifies the paid OpenRouter IDs without :free. This guide preserves those IDs; it does not substitute a free variant or independently certify the runtime configuration.
Not on that evidence alone. NVIDIA describes an underlying model capability, while the website's audited hosting reference and each endpoint have their own limits. Leave room for output and reasoning, and verify the actual served route.
No. A model can produce tool requests, but a compatible, authorized integration must execute them and return results. This website guide does not establish that those tools or settings are available in your app session.
New to the terminology?
Official documentation
Highlighted provider contracts reviewed . Individual model guides carry their own review dates. Sources can change; these pages do not perform live health checks.
NVIDIA's official cards and OpenRouter endpoint documentation support the model identities, visual capabilities and hosting limits below. The four earlier mappings retain their existing owner confirmation. Host-specific reference limits and prices are not universal allowances or measured generation results.