> ## 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.
# Generative models
> Compare the base and instruction-tuned models available for agents, fine-tuning, and inference.
export const SupportedOn = ({ui = false, api = true, sdk = true, sdkVersion = null}) =>
Supported on
UI
API
SDK
{sdk && sdkVersion &&
Requires seekrai {sdkVersion}
}
;
A generative model produces text as output. These models provide the reasoning capabilities needed to understand tasks, determine a course of action, and generate responses. A generative model serves as an agent's cognitive engine, and every agent requires one to function.
Most generative models take text as input. Vision models also accept images, which they describe or reason about in their text response.
Generative models are also the models you fine-tune and deploy for inference. For models that produce vectors rather than text, see [Embedding models](/flow/components/models/embedding-models).
## Model types
### Instruct models
Instruct models are instruction-tuned and ready to use. They follow instructions, hold conversations, and respond to prompts without further training. Most fine-tuning also starts from an Instruct model, since fine-tuning adapts a model that already follows instructions.
### Base models
Base models are pretrained weights without instruction tuning. They are not conversational without further training.
### Fine-tuned models
Fine-tuned models are custom-trained models adapted to specific domains or use cases. These models are created through SeekrFlow's fine-tuning component, which embeds specialized knowledge and behaviors directly into model parameters.
## Available models
The **Type** column reports whether a model is Instruct or Base. Whether a given model can be fine-tuned or deployed is reported per model in the catalog and is not implied by its type. See [Model attributes](/flow/components/models#model-attributes).
### Meta Llama models
| Model | Type | Size | Best for | Country of origin |
| ------------------------------ | -------- | ----------------- | --------------------------------------------------------------- | ----------------- |
| Llama 3.1 8B Instruct | Instruct | 8B | Instruction-following, general assistant tasks | United States |
| Llama 3.1 8B | Base | 8B | Fine-tuning on general text tasks | United States |
| Llama 3.1 70B Instruct | Instruct | 70B | Complex reasoning, high-accuracy tasks | United States |
| Llama 3.1 70B | Base | 70B | Fine-tuning at scale | United States |
| Llama 3.2 1B/3B Instruct | Instruct | 1-3B | Lightweight tasks, resource-constrained environments | United States |
| Llama 3.2 1B/3B | Base | 1-3B | Fine-tuning for multilingual text generation across 8 languages | United States |
| Llama 3.2 11B Vision Instruct | Instruct | 11B | Visual recognition, image reasoning, captioning | United States |
| Llama 3.2 90B Vision Instruct | Instruct | 90B | Multimodal image reasoning and visual Q\&A | United States |
| Llama 3.3 70B Instruct | Instruct | 70B | Latest generation complex reasoning | United States |
| Llama 4 Scout 17B-16E Instruct | Instruct | 109B (17B active) | Next-generation multimodal understanding | United States |
| Meta Llama 3 8B Instruct | Instruct | 8B | Instruction-following on the previous generation | United States |
### Qwen models
| Model | Type | Size | Best for | Country of origin |
| -------------------------- | -------- | ----------------- | ------------------------------------------------------------------------ | ----------------- |
| Qwen 2 72B | Base | 72B | Advanced natural language tasks | China |
| Qwen 2.5 3B | Base | 3B | Fine-tuning for multilingual text generation | China |
| Qwen 2.5 3B Instruct | Instruct | 3B | Multilingual instruction-following | China |
| Qwen 2.5 32B Instruct | Instruct | 32B | Advanced multilingual instruction-following | China |
| Qwen 2.5 72B Instruct | Instruct | 72B | Large-scale multilingual instruction-following | China |
| Qwen 2.5 VL 3B Instruct | Instruct | 3B | Visual question answering, document comprehension, captioning | China |
| Qwen 2.5 VL 7B Instruct | Instruct | 7B | Document and form parsing, OCR, and video understanding | China |
| Qwen 2.5 VL 32B Instruct | Instruct | 32B | Vision and language understanding | China |
| Qwen 3 8B | Instruct | 8B | Efficient multilingual reasoning | China |
| Qwen 3 8B/32B FP8 | Base | 8-32B | Fine-tuning with reduced memory requirements | China |
| Qwen 3 30B A3B FP8 | Base | 30B (3B active) | Balancing performance and efficiency, 128K context | China |
| Qwen 3 30B A3B Thinking | Instruct | 30B (3B active) | Extended reasoning across mathematics, science, and coding, 256K context | China |
| Qwen 3 235B A22B FP8 | Base | 235B (22B active) | Complex tasks at large scale, 128K context | China |
| Qwen 3 Coder 480B A35B FP8 | Instruct | 480B (35B active) | Agentic coding and repository-scale code understanding, 256K context | China |
| Qwen 3 VL 8B Instruct | Instruct | 8B | Visual agent tasks, visual coding, and OCR in 32 languages | China |
| Qwen 3 VL 30B A3B Instruct | Instruct | 30B (3B active) | Visual agents, spatial reasoning, and long video understanding | China |
| Qwen 3.5 9B | Instruct | 9B | Multimodal reasoning, with thinking mode enabled by default | China |
| Qwen 3.5 397B A17B | Instruct | 397B (17B active) | Multimodal reasoning at frontier scale | China |
| Qwen 3.6 27B | Instruct | 27B | Agentic coding and frontend development | China |
### Mistral AI models
| Model | Type | Size | Best for | Country of origin |
| ----------------------------- | -------- | ---- | -------------------------------------- | ----------------- |
| Mistral 7B Instruct v0.2/v0.3 | Instruct | 7B | Fast responses, chat applications | France |
| Mistral Small 24B Instruct | Instruct | 24B | Low-latency multilingual conversations | France |
### DeepSeek models
| Model | Type | Size | Best for | Country of origin |
| ---------------------------- | -------- | ---- | -------------------------------------- | ----------------- |
| DeepSeek R1 Distill Qwen 7B | Instruct | 7B | Reasoning-focused conversational tasks | China |
| DeepSeek R1 Distill Qwen 32B | Instruct | 32B | Advanced reasoning at scale | China |
### Google Gemma models
| Model | Type | Size | Best for | Country of origin |
| ------------------ | -------- | --------------- | ---------------------------------------------------- | ----------------- |
| Gemma 2B | Base | 2B | Fine-tuning for lightweight general tasks | Mixed |
| Gemma 2 9B | Base | 9B | Fine-tuning with balanced performance and efficiency | Mixed |
| Gemma 3 27B IT | Instruct | 27B | Multilingual, multimodal capabilities | Mixed |
| Gemma 4 E2B IT | Instruct | E2B | Text and image input, 256K context, 140+ languages | Mixed |
| Gemma 4 E4B IT | Instruct | E4B | Text and image input, 256K context, 140+ languages | Mixed |
| Gemma 4 26B A4B IT | Instruct | 26B (4B active) | Text and image input, 256K context, 140+ languages | Mixed |
| Gemma 4 31B IT | Instruct | 31B | Text and image input, 256K context, 140+ languages | Mixed |
| Gemma 4 31B IT FP8 | Instruct | 31B (quantized) | Lower memory requirements for inference | Mixed |
### Seekr models
| Model | Type | Size | Best for | Country of origin |
| -------------------------- | -------- | ---- | ---------------------------------------------------------------------------------------------------------------------- | ----------------- |
| Geo 31B Instruct | Instruct | 31B | Satellite imagery and geospatial tasks, including scene description, object identification, and change detection | United States |
| Geo 31B Reasoning | Instruct | 31B | Geospatial reasoning, including object counting, positioning, and hypothesis testing | United States |
| Retail Cosmos 2B Reasoning | Instruct | 2B | Physical scene understanding in retail, including activity recognition, hand-object interaction, and spatial reasoning | United States |
### Other models
| Model | Type | Size | Best for | Country of origin |
| -------------------------------- | -------- | ---------------- | ----------------------------------------------------------------- | ----------------- |
| Microsoft Phi-3 Mini 4k Instruct | Instruct | 3.8B | On-device and resource-constrained environments | United States |
| TinyLlama 1.1B Chat | Instruct | 1.1B | Ultra-lightweight chat applications | Singapore |
| OpenAI GPT-OSS 20B/120B | Instruct | 20-120B | Reasoning and agentic tasks | United States |
| Nous Hermes 2 Mixtral 8x7B DPO | Instruct | 46.7B | Preference-aligned instruction-following | United States |
| Inkling | Instruct | | General-purpose multimodal work with text, image, and audio input | United States |
| Inkling Small | Instruct | | General-purpose multimodal work at a smaller scale | United States |
| Laguna S 2.1 | Base | 118B (8B active) | Agentic coding and long-horizon work | Mixed |
**Mixed** means a model has no single country of origin. Empty cells are values the model catalog does not currently report. See [Country of origin](/flow/components/models#country-of-origin).
Models are added and updated over time. To retrieve the current models and their attributes, see [List models](/flow/reference/ml_models_v1_flow_models_get).
## Select a model
When configuring an agent, you specify which model to use. Consider these factors when selecting a model:
* **Type** – Instruct models are ready to use as-is. Base models require fine-tuning before they behave conversationally.
* **Task complexity** – More complex reasoning tasks may benefit from larger or specialized models.
* **Response requirements** – Balance between speed and accuracy based on your use case.
* **Domain specificity** – Fine-tuned models perform better for specialized domains with unique terminology or requirements.
* **Cost and performance** – Larger models provide enhanced capabilities but with higher computational costs.
* **Country of origin** – Required if your deployment restricts model provenance. See [Country of origin](/flow/components/models#country-of-origin).
## Model configuration
Models are specified during agent creation and can be updated by modifying the agent configuration. The model works in conjunction with other agent components like instructions and tools to determine overall agent behavior.