Nemotron 3 Embed 1B is an NVIDIA text model for semantic retrieval and passage matching. It encodes questions, documents and program code into dense numerical vectors. These representations are used to find relevant context in large collections, add it to an assistant's response or retrieve previously saved information for an agent's next step.
The model is intended for multilingual and cross-language retrieval: a question and relevant material can be written in different languages. Russian belongs to the published set of 34 languages, alongside English, Ukrainian, Chinese, Japanese, Arabic and others. This is useful for shared knowledge bases where technical documents and user requests have not been standardized to one language.
Tasks include documentation search, text matching and code retrieval from a verbal description. A large input window allows work with long documents, extensive program context and the history of several agent steps. Retrieved material returns to the subsequent workflow as context; the next system component produces final explanations and answers.
Input consists of text strings. Each string yields one vector of 2048 numbers. Maximum sequence length is 32,768 tokens; longer material needs to be divided into parts or shortened. Input preparation distinguishes a question from a document: the model was trained to match their roles, so the same string in these two tasks should not be formatted without accounting for its purpose.
Vectors are normalized to unit length. With this representation, dot product and cosine similarity provide equivalent comparison methods. The first 1024 or 512 coordinates of the full vector can be retained, but normalization must be performed again after truncation. This is a result-processing property; dimensionality should be selected consistently for the collection and queries.
The encoder contains approximately 1.14 billion parameters. It is derived from a model based on Ministral 3 3B through two stages of structured pruning and distillation. Bidirectional attention considers the entire input, then token representations are averaged into a shared vector.
NVIDIA releases BF16 and NVFP4 variants. Their representation spaces are close, but switching variants requires checks on one's own documents. Neither numerical similarity nor context length guarantees that retrieval always selects the correct passage.
", "aboutText": "Nemotron 3 Embed 1B is an NVIDIA text model for semantic retrieval and passage matching. It encodes questions, documents and program code into dense numerical vectors. These representations are used to find relevant context in large collections, add it to an assistant's response or retrieve previously saved information for an agent's next step.\n\n### Multilingual knowledge and code\n\nThe model is intended for multilingual and cross-language retrieval: a question and relevant material can be written in different languages. Russian belongs to the published set of 34 languages, alongside English, Ukrainian, Chinese, Japanese, Arabic and others. This is useful for shared knowledge bases where technical documents and user requests have not been standardized to one language.\n\nTasks include documentation search, text matching and code retrieval from a verbal description. A large input window allows work with long documents, extensive program context and the history of several agent steps. Retrieved material returns to the subsequent workflow as context; the next system component produces final explanations and answers.\n\n### From text to a vector\n\nInput consists of text strings. Each string yields one vector of 2048 numbers. Maximum sequence length is 32,768 tokens; longer material needs to be divided into parts or shortened. Input preparation distinguishes a question from a document: the model was trained to match their roles, so the same string in these two tasks should not be formatted without accounting for its purpose.\n\nVectors are normalized to unit length. With this representation, dot product and cosine similarity provide equivalent comparison methods. The first 1024 or 512 coordinates of the full vector can be retained, but normalization must be performed again after truncation. This is a result-processing property; dimensionality should be selected consistently for the collection and queries.\n\n### How the compact model is built\n\nThe encoder contains approximately 1.14 billion parameters. It is derived from a model based on Ministral 3 3B through two stages of structured pruning and distillation. Bidirectional attention considers the entire input, then token representations are averaged into a shared vector.\n\nNVIDIA releases BF16 and NVFP4 variants. Their representation spaces are close, but switching variants requires checks on one's own documents. Neither numerical similarity nor context length guarantees that retrieval always selects the correct passage.", "editorialOverlay": { "copy_fields": null, "copy_sha256": null, "copy_source": null, "html_sha256": "3f7c16e0a7193445f10babfca58d76409e63425f3cfb90a0a916d445e940134c", "module_sha256": "ee9dc1c50b1e6a8fb7fa062030fcb4646b9aaaeb08be513b723a5a92a7290f06", "raw_model_unchanged": true, "schema": "ailoot-native-model-editorial-v1" }, "locale": "en", "methods": [ { "html": "nvidia/nemotron-3-embed-1bNemotron 3 Embed 1B turns text into vectors for multilingual retrieval, code search and RAG. Processes sequences of up to 32,768 tokens and returns a 2048-dimensional representation.
This description covers the model’s capabilities. Available in Ailoot inputs, outputs and options are listed for the selected method below.
Nemotron 3 Embed 1B is an NVIDIA text model for semantic retrieval and passage matching. It encodes questions, documents and program code into dense numerical vectors. These representations are used to find relevant context in large collections, add it to an assistant's response or retrieve previously saved information for an agent's next step.
The model is intended for multilingual and cross-language retrieval: a question and relevant material can be written in different languages. Russian belongs to the published set of 34 languages, alongside English, Ukrainian, Chinese, Japanese, Arabic and others. This is useful for shared knowledge bases where technical documents and user requests have not been standardized to one language.
Tasks include documentation search, text matching and code retrieval from a verbal description. A large input window allows work with long documents, extensive program context and the history of several agent steps. Retrieved material returns to the subsequent workflow as context; the next system component produces final explanations and answers.
Input consists of text strings. Each string yields one vector of 2048 numbers. Maximum sequence length is 32,768 tokens; longer material needs to be divided into parts or shortened. Input preparation distinguishes a question from a document: the model was trained to match their roles, so the same string in these two tasks should not be formatted without accounting for its purpose.
Vectors are normalized to unit length. With this representation, dot product and cosine similarity provide equivalent comparison methods. The first 1024 or 512 coordinates of the full vector can be retained, but normalization must be performed again after truncation. This is a result-processing property; dimensionality should be selected consistently for the collection and queries.
The encoder contains approximately 1.14 billion parameters. It is derived from a model based on Ministral 3 3B through two stages of structured pruning and distillation. Bidirectional attention considers the entire input, then token representations are averaged into a shared vector.
NVIDIA releases BF16 and NVFP4 variants. Their representation spaces are close, but switching variants requires checks on one's own documents. Neither numerical similarity nor context length guarantees that retrieval always selects the correct passage.
POST /v1/embeddingstextembedding| Field | Type | Requirement | Values & limits | Default |
|---|---|---|---|---|
Fieldmodel | Typestring | RequirementRequired | Values & limits nvidia/nemotron-3-embed-1b | Default— |
Fieldinput | Typestring|string[] | RequirementRequired | Values & limits Minimum items: 1 Conditions: Every string nonempty after trim; data URIs are rejected; accepted budget measures relevant UTF8 bytes. | Default— |
Fieldinput_type | Typestring | RequirementOptional | Values & limits query, passage | Defaultquery |
Fieldencoding_format | Typestring | RequirementOptional | Values & limits float, base64 | Defaultfloat |
Fieldtruncate | Typestring | RequirementOptional | Values & limits NONE, START, END | Default— |
Fielddimensions | Typeinteger | RequirementOptional | Values & limits Minimum: 1 | Default— |
Fielduser | Typestring | RequirementOptional | Values & limits | Default— |
text_to_embedding
| Parameter | Contract |
|---|---|
Parametermaterial_types | Contracttext |
Parameterrequired | Contractmodel, input |
Parameteroptional_controls | Contractnull |
Parametercontent_type | Contractapplication/json |
Lifecycle: exact retained published method
Method shapes and IDs are exact; supplier capability is separate from admission.
POST /v1/embeddingsAiloot 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: "${AILOOT_API_KEY:?Set your Ailoot API key}"\n: "${AILOOT_TEXT_FILE:?Set your actual UTF-8 document}"\nexport AILOOT_TEXT_FILE\npython3 - <<'PY_REQUEST'\nimport json,os\nfrom pathlib import Path\np=json.loads('{"model": "nvidia/nemotron-3-embed-1b", "input": []}')\nf=Path(os.environ["AILOOT_TEXT_FILE"])\nif not f.is_file():raise SystemExit("Choose your actual UTF-8 source material")\ntext=f.read_text(encoding="utf-8")\nif not text.strip():raise SystemExit("Source material must contain text")\np["input"]=[text]\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/embeddings \\\n -H "Authorization: Bearer $AILOOT_API_KEY" -H 'Content-Type: application/json' \\\n --data-binary @request.json -D response-headers.txt -o result.json\npython3 - <<'PY_RESULT'\nimport json\nfrom pathlib import Path\nr=json.loads(Path("result.json").read_text())\nfor x in r.get("data",[]): print(json.dumps({"index":x.get("index"),"embedding":x.get("embedding")}))\nPY_RESULT\njson
Keep every complete returned response, file and content part.
Decode base64 vectors only when that format was requested; retain every vector and its original index.
No embedding job endpoint is introduced.
| Field | Type |
|---|---|
Fielddata[].index | Typeinteger |
Fielddata[].embedding | Typefinite float[]|base64 float32 vector |
Fieldmodel | Typestring |
Fieldusage | Typeobject |
| Response | Condition | Next step |
|---|---|---|
Response400 · invalid_request_error | ConditionInvalid required fields, material representation or parameter value | Next stepCorrect the indicated fields before a new creation request. |
Ailoot nvidia/nemotron-3-embed-1b, nemotron-3-embed, nemotron-3-embed-latest
Use the method paths shown above.
nvidia/nemotron-3-embed-1b, nemotron-3-embed, nemotron-3-embed-latest
nemotron-3-embedNemotron 3 Embed 1B turns text into vectors for multilingual retrieval, code search and RAG. Processes sequences of up to 32,768 tokens and returns a 2048-dimensional representation.
This description covers the model’s capabilities. Available in Ailoot inputs, outputs and options are listed for the selected method below.
Nemotron 3 Embed 1B is an NVIDIA text model for semantic retrieval and passage matching. It encodes questions, documents and program code into dense numerical vectors. These representations are used to find relevant context in large collections, add it to an assistant's response or retrieve previously saved information for an agent's next step.
The model is intended for multilingual and cross-language retrieval: a question and relevant material can be written in different languages. Russian belongs to the published set of 34 languages, alongside English, Ukrainian, Chinese, Japanese, Arabic and others. This is useful for shared knowledge bases where technical documents and user requests have not been standardized to one language.
Tasks include documentation search, text matching and code retrieval from a verbal description. A large input window allows work with long documents, extensive program context and the history of several agent steps. Retrieved material returns to the subsequent workflow as context; the next system component produces final explanations and answers.
Input consists of text strings. Each string yields one vector of 2048 numbers. Maximum sequence length is 32,768 tokens; longer material needs to be divided into parts or shortened. Input preparation distinguishes a question from a document: the model was trained to match their roles, so the same string in these two tasks should not be formatted without accounting for its purpose.
Vectors are normalized to unit length. With this representation, dot product and cosine similarity provide equivalent comparison methods. The first 1024 or 512 coordinates of the full vector can be retained, but normalization must be performed again after truncation. This is a result-processing property; dimensionality should be selected consistently for the collection and queries.
The encoder contains approximately 1.14 billion parameters. It is derived from a model based on Ministral 3 3B through two stages of structured pruning and distillation. Bidirectional attention considers the entire input, then token representations are averaged into a shared vector.
NVIDIA releases BF16 and NVFP4 variants. Their representation spaces are close, but switching variants requires checks on one's own documents. Neither numerical similarity nor context length guarantees that retrieval always selects the correct passage.
POST /v1/embeddingstextembedding| Field | Type | Requirement | Values & limits | Default |
|---|---|---|---|---|
Fieldmodel | Typestring | RequirementRequired | Values & limits nemotron-3-embed | Default— |
Fieldinput | Typestring|string[] | RequirementRequired | Values & limits Minimum items: 1 Conditions: Every string nonempty after trim; data URIs are rejected; accepted budget measures relevant UTF8 bytes. | Default— |
Fieldinput_type | Typestring | RequirementOptional | Values & limits query, passage | Defaultquery |
Fieldencoding_format | Typestring | RequirementOptional | Values & limits float, base64 | Defaultfloat |
Fieldtruncate | Typestring | RequirementOptional | Values & limits NONE, START, END | Default— |
Fielddimensions | Typeinteger | RequirementOptional | Values & limits Minimum: 1 | Default— |
Fielduser | Typestring | RequirementOptional | Values & limits | Default— |
text_to_embedding
| Parameter | Contract |
|---|---|
Parametermaterial_types | Contracttext |
Parameterrequired | Contractmodel, input |
Parameteroptional_controls | Contractnull |
Parametercontent_type | Contractapplication/json |
Lifecycle: exact retained published method
Method shapes and IDs are exact; supplier capability is separate from admission.
POST /v1/embeddingsAiloot 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}"\n: "${AILOOT_TEXT_FILE:?Set your actual UTF-8 document}"\nexport AILOOT_TEXT_FILE\npython3 - <<'PY_REQUEST'\nimport json,os\nfrom pathlib import Path\np=json.loads('{"model": "nemotron-3-embed", "input": []}')\nf=Path(os.environ["AILOOT_TEXT_FILE"])\nif not f.is_file():raise SystemExit("Choose your actual UTF-8 source material")\ntext=f.read_text(encoding="utf-8")\nif not text.strip():raise SystemExit("Source material must contain text")\np["input"]=[text]\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/embeddings \\\n -H "Authorization: Bearer $AILOOT_API_KEY" -H 'Content-Type: application/json' \\\n --data-binary @request.json -D response-headers.txt -o result.json\npython3 - <<'PY_RESULT'\nimport json\nfrom pathlib import Path\nr=json.loads(Path("result.json").read_text())\nfor x in r.get("data",[]): print(json.dumps({"index":x.get("index"),"embedding":x.get("embedding")}))\nPY_RESULT\njson
Keep every complete returned response, file and content part.
Decode base64 vectors only when that format was requested; retain every vector and its original index.
No embedding job endpoint is introduced.
| Field | Type |
|---|---|
Fielddata[].index | Typeinteger |
Fielddata[].embedding | Typefinite float[]|base64 float32 vector |
Fieldmodel | Typestring |
Fieldusage | Typeobject |
| Response | Condition | Next step |
|---|---|---|
Response400 · invalid_request_error | ConditionInvalid required fields, material representation or parameter value | Next stepCorrect the indicated fields before a new creation request. |
Ailoot nvidia/nemotron-3-embed-1b, nemotron-3-embed, nemotron-3-embed-latest
Use the method paths shown above.
nvidia/nemotron-3-embed-1b, nemotron-3-embed, nemotron-3-embed-latest
nemotron-3-embed-latestNemotron 3 Embed 1B turns text into vectors for multilingual retrieval, code search and RAG. Processes sequences of up to 32,768 tokens and returns a 2048-dimensional representation.
This description covers the model’s capabilities. Available in Ailoot inputs, outputs and options are listed for the selected method below.
Nemotron 3 Embed 1B is an NVIDIA text model for semantic retrieval and passage matching. It encodes questions, documents and program code into dense numerical vectors. These representations are used to find relevant context in large collections, add it to an assistant's response or retrieve previously saved information for an agent's next step.
The model is intended for multilingual and cross-language retrieval: a question and relevant material can be written in different languages. Russian belongs to the published set of 34 languages, alongside English, Ukrainian, Chinese, Japanese, Arabic and others. This is useful for shared knowledge bases where technical documents and user requests have not been standardized to one language.
Tasks include documentation search, text matching and code retrieval from a verbal description. A large input window allows work with long documents, extensive program context and the history of several agent steps. Retrieved material returns to the subsequent workflow as context; the next system component produces final explanations and answers.
Input consists of text strings. Each string yields one vector of 2048 numbers. Maximum sequence length is 32,768 tokens; longer material needs to be divided into parts or shortened. Input preparation distinguishes a question from a document: the model was trained to match their roles, so the same string in these two tasks should not be formatted without accounting for its purpose.
Vectors are normalized to unit length. With this representation, dot product and cosine similarity provide equivalent comparison methods. The first 1024 or 512 coordinates of the full vector can be retained, but normalization must be performed again after truncation. This is a result-processing property; dimensionality should be selected consistently for the collection and queries.
The encoder contains approximately 1.14 billion parameters. It is derived from a model based on Ministral 3 3B through two stages of structured pruning and distillation. Bidirectional attention considers the entire input, then token representations are averaged into a shared vector.
NVIDIA releases BF16 and NVFP4 variants. Their representation spaces are close, but switching variants requires checks on one's own documents. Neither numerical similarity nor context length guarantees that retrieval always selects the correct passage.
POST /v1/embeddingstextembedding| Field | Type | Requirement | Values & limits | Default |
|---|---|---|---|---|
Fieldmodel | Typestring | RequirementRequired | Values & limits nemotron-3-embed-latest | Default— |
Fieldinput | Typestring|string[] | RequirementRequired | Values & limits Minimum items: 1 Conditions: Every string nonempty after trim; data URIs are rejected; accepted budget measures relevant UTF8 bytes. | Default— |
Fieldinput_type | Typestring | RequirementOptional | Values & limits query, passage | Defaultquery |
Fieldencoding_format | Typestring | RequirementOptional | Values & limits float, base64 | Defaultfloat |
Fieldtruncate | Typestring | RequirementOptional | Values & limits NONE, START, END | Default— |
Fielddimensions | Typeinteger | RequirementOptional | Values & limits Minimum: 1 | Default— |
Fielduser | Typestring | RequirementOptional | Values & limits | Default— |
text_to_embedding
| Parameter | Contract |
|---|---|
Parametermaterial_types | Contracttext |
Parameterrequired | Contractmodel, input |
Parameteroptional_controls | Contractnull |
Parametercontent_type | Contractapplication/json |
Lifecycle: exact retained published method
Method shapes and IDs are exact; supplier capability is separate from admission.
POST /v1/embeddingsAiloot 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}"\n: "${AILOOT_TEXT_FILE:?Set your actual UTF-8 document}"\nexport AILOOT_TEXT_FILE\npython3 - <<'PY_REQUEST'\nimport json,os\nfrom pathlib import Path\np=json.loads('{"model": "nemotron-3-embed-latest", "input": []}')\nf=Path(os.environ["AILOOT_TEXT_FILE"])\nif not f.is_file():raise SystemExit("Choose your actual UTF-8 source material")\ntext=f.read_text(encoding="utf-8")\nif not text.strip():raise SystemExit("Source material must contain text")\np["input"]=[text]\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/embeddings \\\n -H "Authorization: Bearer $AILOOT_API_KEY" -H 'Content-Type: application/json' \\\n --data-binary @request.json -D response-headers.txt -o result.json\npython3 - <<'PY_RESULT'\nimport json\nfrom pathlib import Path\nr=json.loads(Path("result.json").read_text())\nfor x in r.get("data",[]): print(json.dumps({"index":x.get("index"),"embedding":x.get("embedding")}))\nPY_RESULT\njson
Keep every complete returned response, file and content part.
Decode base64 vectors only when that format was requested; retain every vector and its original index.
No embedding job endpoint is introduced.
| Field | Type |
|---|---|
Fielddata[].index | Typeinteger |
Fielddata[].embedding | Typefinite float[]|base64 float32 vector |
Fieldmodel | Typestring |
Fieldusage | Typeobject |
| Response | Condition | Next step |
|---|---|---|
Response400 · invalid_request_error | ConditionInvalid required fields, material representation or parameter value | Next stepCorrect the indicated fields before a new creation request. |
Ailoot nvidia/nemotron-3-embed-1b, nemotron-3-embed, nemotron-3-embed-latest
Use the method paths shown above.
nvidia/nemotron-3-embed-1b, nemotron-3-embed, nemotron-3-embed-latest