NVIDIA Nemotron 3.5 Lightning is a language model for reasoning, conversation and specialized agent workflows. It has 30B total parameters with 3B active, using a hybrid architecture that interleaves Mamba-2, mixture-of-experts and attention layers. Multi-Token Prediction layers add training signals for future tokens and support the model’s optimized generation recipes. The model processes text and produces text, including explanations, code and answers grounded in supplied material.
The manufacturer describes applications in chatbots, retrieval-augmented generation, instruction following and agents. For a knowledge-base answer, provide the question and the relevant passages with stable source labels. Ask the model to answer from those passages, identify which one supports each important claim, and state what remains unknown. For example: “Using these support articles, explain the difference between the two installation methods and return the prerequisites, steps and checks for each.” Keep a missing prerequisite visible instead of inviting the model to invent it.
For coding work, supply the relevant code, expected behavior and observed failure. A practical request asks for a focused revision and an explanation of the affected cases. If a tool-connected agent runs a check, return its actual output before asking for the next change. The model’s training includes code, multi-step tool use and structured-output tasks, but the completed artifact still needs to match the original requirement. Ask for the revised code, the reason for each material change and the observations needed to establish that the problem is resolved.
The maker documents thinking enabled by default and a direct-answer mode with thinking disabled through the chat template. Reasoning mode is useful for a task with several dependent conditions; a direct answer can suit straightforward transformation or extraction. The manufacturer recommends temperature 1.0 and top-p 0.95 in its client examples. Treat those as maker recipe settings and choose the final response structure explicitly, such as a concise explanation, a comparison table or an answer followed by its supporting evidence.
The model supports up to a 1M-token context. The published BF16 deployment recipes also distinguish that model capability from configurations using a 256K window under memory constraints. Prepare long material for the applicable working window, preserving the document structure and leaving room for the response. A long context is most useful when the prompt specifies the relationship to find: contradictions between documents, dependencies across source files, or the evidence supporting a conclusion. It is not a reason to include unrelated history or to assume a final answer can occupy the entire window.
English and coding languages are the primary focus, with Spanish, French, German, Italian and Japanese also listed. State the required output language and retain names, identifiers and quoted source terms when their spelling matters. For multilingual document work, specify whether to translate, summarize or compare the materials; these requests produce different artifacts. Ask for uncertain terminology to be flagged so a fluent translation does not silently change a technical meaning.
NVIDIA publishes a BF16 reference release primarily for customization and full-precision research, and a separate NVFP4 release with optimized inference recipes. This distinction matters when comparing deployment guidance: a hardware configuration or throughput result from one release does not automatically describe another. The model name identifies the Lightning family, while the precise checkpoint and serving configuration determine the recipe. Keep model-level capabilities separate from recipe-specific performance and memory assumptions.
For tool-based work, provide clear function descriptions and return the actual result of each executed step. A proposed operation becomes usable evidence only after the surrounding application carries it out and records what happened. Ask the final answer to distinguish source observations, derived conclusions and unfinished checks. When a calculation, command or retrieved passage conflicts with the planned result, use that conflict to revise the conclusion rather than producing a polished summary of an unverified assumption.
", "aboutText": "NVIDIA Nemotron 3.5 Lightning is a language model for reasoning, conversation and specialized agent workflows. It has 30B total parameters with 3B active, using a hybrid architecture that interleaves Mamba-2, mixture-of-experts and attention layers. Multi-Token Prediction layers add training signals for future tokens and support the model’s optimized generation recipes. The model processes text and produces text, including explanations, code and answers grounded in supplied material.\n\nThe manufacturer describes applications in chatbots, retrieval-augmented generation, instruction following and agents. For a knowledge-base answer, provide the question and the relevant passages with stable source labels. Ask the model to answer from those passages, identify which one supports each important claim, and state what remains unknown. For example: “Using these support articles, explain the difference between the two installation methods and return the prerequisites, steps and checks for each.” Keep a missing prerequisite visible instead of inviting the model to invent it.\n\nFor coding work, supply the relevant code, expected behavior and observed failure. A practical request asks for a focused revision and an explanation of the affected cases. If a tool-connected agent runs a check, return its actual output before asking for the next change. The model’s training includes code, multi-step tool use and structured-output tasks, but the completed artifact still needs to match the original requirement. Ask for the revised code, the reason for each material change and the observations needed to establish that the problem is resolved.\n\nThe maker documents thinking enabled by default and a direct-answer mode with thinking disabled through the chat template. Reasoning mode is useful for a task with several dependent conditions; a direct answer can suit straightforward transformation or extraction. The manufacturer recommends temperature 1.0 and top-p 0.95 in its client examples. Treat those as maker recipe settings and choose the final response structure explicitly, such as a concise explanation, a comparison table or an answer followed by its supporting evidence.\n\nThe model supports up to a 1M-token context. The published BF16 deployment recipes also distinguish that model capability from configurations using a 256K window under memory constraints. Prepare long material for the applicable working window, preserving the document structure and leaving room for the response. A long context is most useful when the prompt specifies the relationship to find: contradictions between documents, dependencies across source files, or the evidence supporting a conclusion. It is not a reason to include unrelated history or to assume a final answer can occupy the entire window.\n\nEnglish and coding languages are the primary focus, with Spanish, French, German, Italian and Japanese also listed. State the required output language and retain names, identifiers and quoted source terms when their spelling matters. For multilingual document work, specify whether to translate, summarize or compare the materials; these requests produce different artifacts. Ask for uncertain terminology to be flagged so a fluent translation does not silently change a technical meaning.\n\nNVIDIA publishes a BF16 reference release primarily for customization and full-precision research, and a separate NVFP4 release with optimized inference recipes. This distinction matters when comparing deployment guidance: a hardware configuration or throughput result from one release does not automatically describe another. The model name identifies the Lightning family, while the precise checkpoint and serving configuration determine the recipe. Keep model-level capabilities separate from recipe-specific performance and memory assumptions.\n\nFor tool-based work, provide clear function descriptions and return the actual result of each executed step. A proposed operation becomes usable evidence only after the surrounding application carries it out and records what happened. Ask the final answer to distinguish source observations, derived conclusions and unfinished checks. When a calculation, command or retrieved passage conflicts with the planned result, use that conflict to revise the conclusion rather than producing a polished summary of an unverified assumption.", "editorialOverlay": { "copy_fields": null, "copy_sha256": null, "copy_source": null, "html_sha256": "132d2f7668c6a4ee19d4134c9e17cf381dfbe443e4bceb21fbac2fb0e67e5b93", "module_sha256": "ee9dc1c50b1e6a8fb7fa062030fcb4646b9aaaeb08be513b723a5a92a7290f06", "raw_model_unchanged": true, "schema": "ailoot-native-model-editorial-v1" }, "locale": "en", "methods": [ { "html": "nvidia/nemotron-3.5-lightning:freeNVIDIA Nemotron 3.5 Lightning is a text reasoning and chat model with 30B total and 3B active parameters. Its hybrid Mamba–MoE architecture supports long-context analysis, coding and tool-based workflows, with thinking or direct-answer modes documented in the maker’s chat template.
This description covers the model’s capabilities. Available in Ailoot inputs, outputs and options are listed for the selected method below.
NVIDIA Nemotron 3.5 Lightning is a language model for reasoning, conversation and specialized agent workflows. It has 30B total parameters with 3B active, using a hybrid architecture that interleaves Mamba-2, mixture-of-experts and attention layers. Multi-Token Prediction layers add training signals for future tokens and support the model’s optimized generation recipes. The model processes text and produces text, including explanations, code and answers grounded in supplied material.
The manufacturer describes applications in chatbots, retrieval-augmented generation, instruction following and agents. For a knowledge-base answer, provide the question and the relevant passages with stable source labels. Ask the model to answer from those passages, identify which one supports each important claim, and state what remains unknown. For example: “Using these support articles, explain the difference between the two installation methods and return the prerequisites, steps and checks for each.” Keep a missing prerequisite visible instead of inviting the model to invent it.
For coding work, supply the relevant code, expected behavior and observed failure. A practical request asks for a focused revision and an explanation of the affected cases. If a tool-connected agent runs a check, return its actual output before asking for the next change. The model’s training includes code, multi-step tool use and structured-output tasks, but the completed artifact still needs to match the original requirement. Ask for the revised code, the reason for each material change and the observations needed to establish that the problem is resolved.
The maker documents thinking enabled by default and a direct-answer mode with thinking disabled through the chat template. Reasoning mode is useful for a task with several dependent conditions; a direct answer can suit straightforward transformation or extraction. The manufacturer recommends temperature 1.0 and top-p 0.95 in its client examples. Treat those as maker recipe settings and choose the final response structure explicitly, such as a concise explanation, a comparison table or an answer followed by its supporting evidence.
The model supports up to a 1M-token context. The published BF16 deployment recipes also distinguish that model capability from configurations using a 256K window under memory constraints. Prepare long material for the applicable working window, preserving the document structure and leaving room for the response. A long context is most useful when the prompt specifies the relationship to find: contradictions between documents, dependencies across source files, or the evidence supporting a conclusion. It is not a reason to include unrelated history or to assume a final answer can occupy the entire window.
English and coding languages are the primary focus, with Spanish, French, German, Italian and Japanese also listed. State the required output language and retain names, identifiers and quoted source terms when their spelling matters. For multilingual document work, specify whether to translate, summarize or compare the materials; these requests produce different artifacts. Ask for uncertain terminology to be flagged so a fluent translation does not silently change a technical meaning.
NVIDIA publishes a BF16 reference release primarily for customization and full-precision research, and a separate NVFP4 release with optimized inference recipes. This distinction matters when comparing deployment guidance: a hardware configuration or throughput result from one release does not automatically describe another. The model name identifies the Lightning family, while the precise checkpoint and serving configuration determine the recipe. Keep model-level capabilities separate from recipe-specific performance and memory assumptions.
For tool-based work, provide clear function descriptions and return the actual result of each executed step. A proposed operation becomes usable evidence only after the surrounding application carries it out and records what happened. Ask the final answer to distinguish source observations, derived conclusions and unfinished checks. When a calculation, command or retrieved passage conflicts with the planned result, use that conflict to revise the conclusion rather than producing a polished summary of an unverified assumption.
POST /v1/chat/completionstexttext| Field | Type | Requirement | Values & limits | Default |
|---|---|---|---|---|
Fieldmodel | Typestring | RequirementRequired | Values & limits nvidia/nemotron-3.5-lightning:free | Default— |
Fieldmessages | Typearray | RequirementRequired | Values & limits | Default— |
Fieldstream | Typeboolean | RequirementOptional | Values & limits false | Defaultfalse |
Fieldmax_tokens | Typeinteger | RequirementOptional | Values & limits Minimum: 1 Maximum: 65,536 | Default— |
Fieldmax_completion_tokens | Typeinteger | RequirementOptional | Values & limits Minimum: 1 Maximum: 65,536 | Default— |
Fieldn | Typeinteger | RequirementOptional | Values & limits Minimum: 1 Maximum: 128 | Default— |
Fieldtemperature | Typenumber | RequirementOptional | Values & limits | Default— |
Fieldtop_p | Typenumber | RequirementOptional | Values & limits | Default— |
Fieldseed | Typeinteger | RequirementOptional | Values & limits | Default— |
Fieldstop | Type | RequirementOptional | Values & limits | Default— |
Fieldfrequency_penalty | Typenumber | RequirementOptional | Values & limits | Default— |
Fieldpresence_penalty | Typenumber | RequirementOptional | Values & limits | Default— |
Fielduser | Typestring | RequirementOptional | Values & limits | Default— |
| Parameter | Contract |
|---|---|
Parametercontent_type | Contractapplication/json |
Parameterrequired | Contractmodel, messages |
Parameteroptional | Contractstream, max_tokens, max_completion_tokens, n, temperature, top_p, seed, stop, frequency_penalty, presence_penalty, user |
Parametermaterial_types | Contracttext |
Parameterconditions | ContractUse literal text messages only. Cache controls, tools, image/audio/video content and streamed responses are unavailable for this method., The complete encoded JSON request must not exceed 1000000 bytes. |
Lifecycle: synchronous_json
Tariff references do not add request options. Supported parameters are listed for the selected method.
| Configuration | Rate schedule | Charge |
|---|---|---|
| ConfigurationBase token rates | Rate scheduleBase rate | Charge Input · Input tokens $0 / 1K tokens Output · Output tokens $0 / 1K tokens Input · Cache read — / 1K tokens Input · Cache write — / 1K tokens Input · Cache write, 1 hour — / 1K tokens |
Ailoot reserves an estimated amount before a request. After the result, it settles the final charge from measured usage. Streaming settlement uses the final usage record. How reservation and settlement work →
set -eu\n: "${AILOOT_API_KEY:?Set your Ailoot API key}"\nexport AILOOT_API_KEY\npython3 - <<'PY_REQUEST'\nimport json,os\nfrom pathlib import Path\np=json.loads('{"model": "nvidia/nemotron-3.5-lightning:free", "messages": [{"role": "user", "content": "Summarize the supplied text in three concise points."}], "stream": false}')\n\nPath("request.json").write_text(json.dumps(p,ensure_ascii=False))\nPY_REQUEST\ncurl --fail-with-body --silent --show-error https://api.ailoot.ai/v1/chat/completions \\\n -H "Authorization: Bearer $AILOOT_API_KEY" -H 'Content-Type: application/json' \\\n --data-binary @request.json -D response-headers.txt -o result.json\n# Submit once. Retain the complete result, original identifiers and every response header.\nJSON
Keep every complete returned response, file and content part.
Retain the complete response, all actual results, original identifiers, usage and response headers.
Preserve the original response, headers and X-Client-Request-ID. An uncertain outcome does not authorize another creation.
Ailoot Ultra nvidia/nemotron-3.5-lightning:free
Use the method paths shown above.
nvidia/nemotron-3.5-lightning:freeNVIDIA Nemotron 3.5 Lightning is a text reasoning and chat model with 30B total and 3B active parameters. Its hybrid Mamba–MoE architecture supports long-context analysis, coding and tool-based workflows, with thinking or direct-answer modes documented in the maker’s chat template.
This description covers the model’s capabilities. Available in Ailoot inputs, outputs and options are listed for the selected method below.
NVIDIA Nemotron 3.5 Lightning is a language model for reasoning, conversation and specialized agent workflows. It has 30B total parameters with 3B active, using a hybrid architecture that interleaves Mamba-2, mixture-of-experts and attention layers. Multi-Token Prediction layers add training signals for future tokens and support the model’s optimized generation recipes. The model processes text and produces text, including explanations, code and answers grounded in supplied material.
The manufacturer describes applications in chatbots, retrieval-augmented generation, instruction following and agents. For a knowledge-base answer, provide the question and the relevant passages with stable source labels. Ask the model to answer from those passages, identify which one supports each important claim, and state what remains unknown. For example: “Using these support articles, explain the difference between the two installation methods and return the prerequisites, steps and checks for each.” Keep a missing prerequisite visible instead of inviting the model to invent it.
For coding work, supply the relevant code, expected behavior and observed failure. A practical request asks for a focused revision and an explanation of the affected cases. If a tool-connected agent runs a check, return its actual output before asking for the next change. The model’s training includes code, multi-step tool use and structured-output tasks, but the completed artifact still needs to match the original requirement. Ask for the revised code, the reason for each material change and the observations needed to establish that the problem is resolved.
The maker documents thinking enabled by default and a direct-answer mode with thinking disabled through the chat template. Reasoning mode is useful for a task with several dependent conditions; a direct answer can suit straightforward transformation or extraction. The manufacturer recommends temperature 1.0 and top-p 0.95 in its client examples. Treat those as maker recipe settings and choose the final response structure explicitly, such as a concise explanation, a comparison table or an answer followed by its supporting evidence.
The model supports up to a 1M-token context. The published BF16 deployment recipes also distinguish that model capability from configurations using a 256K window under memory constraints. Prepare long material for the applicable working window, preserving the document structure and leaving room for the response. A long context is most useful when the prompt specifies the relationship to find: contradictions between documents, dependencies across source files, or the evidence supporting a conclusion. It is not a reason to include unrelated history or to assume a final answer can occupy the entire window.
English and coding languages are the primary focus, with Spanish, French, German, Italian and Japanese also listed. State the required output language and retain names, identifiers and quoted source terms when their spelling matters. For multilingual document work, specify whether to translate, summarize or compare the materials; these requests produce different artifacts. Ask for uncertain terminology to be flagged so a fluent translation does not silently change a technical meaning.
NVIDIA publishes a BF16 reference release primarily for customization and full-precision research, and a separate NVFP4 release with optimized inference recipes. This distinction matters when comparing deployment guidance: a hardware configuration or throughput result from one release does not automatically describe another. The model name identifies the Lightning family, while the precise checkpoint and serving configuration determine the recipe. Keep model-level capabilities separate from recipe-specific performance and memory assumptions.
For tool-based work, provide clear function descriptions and return the actual result of each executed step. A proposed operation becomes usable evidence only after the surrounding application carries it out and records what happened. Ask the final answer to distinguish source observations, derived conclusions and unfinished checks. When a calculation, command or retrieved passage conflicts with the planned result, use that conflict to revise the conclusion rather than producing a polished summary of an unverified assumption.
POST /v1/messagestexttext| Field | Type | Requirement | Values & limits | Default |
|---|---|---|---|---|
Fieldmodel | Typestring | RequirementRequired | Values & limits nvidia/nemotron-3.5-lightning:free | Default— |
Fieldmessages | Typearray | RequirementRequired | Values & limits | Default— |
Fieldstream | Typeboolean | RequirementOptional | Values & limits false | Defaultfalse |
Fieldsystem | Typestring | RequirementOptional | Values & limits | Default— |
Fieldmax_tokens | Typeinteger | RequirementRequired | Values & limits Minimum: 1 Maximum: 65,536 | Default— |
Fieldtemperature | Typenumber | RequirementOptional | Values & limits | Default— |
Fieldtop_p | Typenumber | RequirementOptional | Values & limits | Default— |
Fieldtop_k | Typeinteger | RequirementOptional | Values & limits | Default— |
Fieldstop_sequences | Typearray | RequirementOptional | Values & limits | Default— |
Fieldmetadata | Typeobject | RequirementOptional | Values & limits | Default— |
| Parameter | Contract |
|---|---|
Parametercontent_type | Contractapplication/json |
Parameterrequired | Contractmodel, messages, max_tokens |
Parameteroptional | Contractstream, system, temperature, top_p, top_k, stop_sequences, metadata |
Parametermaterial_types | Contracttext |
Parameterconditions | ContractUse literal text messages only. Cache controls, tools, image/audio/video content and streamed responses are unavailable for this method., The complete encoded JSON request must not exceed 1000000 bytes. |
Lifecycle: synchronous_json
Tariff references do not add request options. Supported parameters are listed for the selected method.
| Configuration | Rate schedule | Charge |
|---|---|---|
| ConfigurationBase token rates | Rate scheduleBase rate | Charge Input · Input tokens $0 / 1K tokens Output · Output tokens $0 / 1K tokens Input · Cache read — / 1K tokens Input · Cache write — / 1K tokens Input · Cache write, 1 hour — / 1K tokens |
Ailoot reserves an estimated amount before a request. After the result, it settles the final charge from measured usage. Streaming settlement uses the final usage record. How reservation and settlement work →
set -eu\n: "${AILOOT_API_KEY:?Set your Ailoot API key}"\nexport AILOOT_API_KEY\npython3 - <<'PY_REQUEST'\nimport json,os\nfrom pathlib import Path\np=json.loads('{"model": "nvidia/nemotron-3.5-lightning:free", "messages": [{"role": "user", "content": "Summarize the supplied text in three concise points."}], "stream": false, "max_tokens": 256}')\n\nPath("request.json").write_text(json.dumps(p,ensure_ascii=False))\nPY_REQUEST\ncurl --fail-with-body --silent --show-error https://api.ailoot.ai/v1/messages \\\n -H "Authorization: Bearer $AILOOT_API_KEY" -H 'Content-Type: application/json' \\\n --data-binary @request.json -D response-headers.txt -o result.json\n# Submit once. Retain the complete result, original identifiers and every response header.\nJSON
Keep every complete returned response, file and content part.
Retain the complete response, all actual results, original identifiers, usage and response headers.
Preserve the original response, headers and X-Client-Request-ID. An uncertain outcome does not authorize another creation.
Ailoot Ultra nvidia/nemotron-3.5-lightning:free
Use the method paths shown above.