> ## Documentation Index
> Fetch the complete documentation index at: https://docs.seekr.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Models
> Understand the types of models available on SeekrFlow and the attributes the catalog reports for each.
export const SupportedOn = ({ui = false, api = true, sdk = true, sdkVersion = null}) =>
Supported on
UI
API
SDK
{sdk && sdkVersion &&
Requires seekrai {sdkVersion}
}
;
The model catalog lists the models available on SeekrFlow, along with each model's identifier, type, capabilities, and country of origin. You select a model from the catalog when you configure an agent, start a fine-tuning job, or create a vector database.
## Model types
The catalog contains two classes of models, distinguished by what they produce.
Produce text as output. These models reason, follow instructions, and generate responses. They serve as an agent's cognitive engine and are the models you fine-tune and deploy for inference.
Produce a fixed-length vector of numbers as output, and generate no language. Embedding models turn documents and queries into comparable vectors, which is what makes semantic search in a vector database possible.
Both classes include models that accept images alongside text. A generative vision model describes or reasons about an image, while a multimodal embedding model represents it as a vector.
The two classes are not interchangeable. A generative model cannot build a vector database, and an embedding model cannot answer a prompt.
## Model attributes
Every model in the catalog reports the following attributes.
| Attribute | Description |
| ------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------ |
| **Name** | The full model identifier, for example `ibm-granite/granite-embedding-311m-multilingual-r2`. Use this value wherever a model is specified. |
| **Model type** | `Instruct`, `Base`, or `Embedding`. See [Model types](#model-types). |
| **Country of origin** | The country the model originated from, or **Mixed** when it has no single country of origin. See [Country of origin](#country-of-origin). |
| **Embedding dimension** | The length of the vector an embedding model produces. Reported for embedding models only. |
| **Inference available** | Whether the model can serve inference requests. |
| **Training available** | Whether the model can be used as the starting point for a fine-tuning job. |
| **Deployment available** | Whether the model can be deployed for inference. See [Deployments](/flow/components/deployments). |
| **Supports LoRA** | Whether the model supports LoRA fine-tuning, and at what maximum rank. |
| **Vision model** | Whether the model accepts image input alongside text. |
### Country of origin
Country of origin identifies where a model originated. Deployments with model provenance requirements, including government and other highly regulated environments, use it to determine whether a model is permitted for a given workload.
The catalog spans several countries of origin. Review this attribute before selecting a model if your deployment restricts model provenance.
Some models have no single country of origin and are listed as **Mixed**. Mixed origins apply when a model's corporate parent and its development lab are based in different countries, or when a model was fine-tuned in one country from a base model developed in another.
Confirm provenance directly with Seekr before using a model with mixed origins in a deployment that restricts model provenance.
## Retrieve the catalog
Retrieve the full catalog, including every attribute, from the models endpoint. Filter by `model_type` to return a single class of models.
See [List models](/flow/reference/ml_models_v1_flow_models_get).
## Select a model
Choosing a generative model and choosing an embedding model involve different trade-offs, and the two decisions carry different consequences.
For generative models, weigh task complexity, response speed, domain specificity, and cost. You can change an agent's model at any time by updating its configuration. See [Generative models](/flow/components/models/generative-models).
For embedding models, weigh vector dimension, maximum input length, and language coverage. This choice is durable. The embedding model is fixed when you create a vector database and cannot be changed afterward without re-embedding your documents. See [Embedding models](/flow/components/models/embedding-models).
If your deployment restricts model provenance, filter by [country of origin](#country-of-origin) before comparing anything else.