Selecting a model
Choosing the right model for your application is a crucial first step. Ai2 offers a diverse range of models, each with unique strengths and capabilities to suit different use cases. Understanding each model's strengths will help you select the optimal solution for your specific requirements, whether you prioritize performance, efficiency, or specialized functionality. For applications requiring maximum throughput with limited resources, OlmoE might be ideal; for document-heavy workflows, olmOCR offers specialized capabilities; and for general assistant-like functionality, Tülu provides robust instruction-following abilities.
Getting started with model selection
When selecting a model, consider these key factors:
- Task requirements: What specific capabilities does your application need?
- Performance needs: How important is accuracy vs. speed for your use case?
- Resource constraints: What are your deployment environment's limitations?
Try before you decide
The easiest way to explore Ai2's models is through the Ai2 Playground, where you can experiment with different models interactively without any setup. This hands-on experience can help you understand each model's strengths and response patterns before integrating them into your application.
Ai2 model family comparison
- Olmo 3
- T ülu 3
- Molmo
- olmOCR
- OlmoE
| Specification | Details |
|---|---|
| Modality | Text |
| Best For | Reasoning, fine-tuning, research |
| Available Sizes | 7B Base, Think, Instruct (14.6GB), 32B Base, Think (64.5GB) |
| License | Apache 2.0 |
The latest generation of Olmo models, pre-trained on the Dolma 3 dataset and post-trained on the Dolci datasets. The Think variants use long chain-of-thought reasoning to improve performance on tasks like math and coding. Fully open with all code, checkpoints, and training details available.
| Specification | Details |
|---|---|
| Modality | Text |
| Best For | Instruction-following assistants (LLaMA 3.1 based) |
| Available Sizes | 8B (16.1GB), 70B (141GB), 405B (810GB est.) |
| License | Meta Llama 3 License + Apache 2.0 (post-training only) |
Built on Meta's LLaMA 3.1 architecture, these models are optimized for instruction-following and assistant-style behavior. With parameter counts ranging from 8B to 405B, developers can choose the right size for their specific performance and resource requirements. The 405B variant is competitive with offerings like GPT-4, while offering more flexibility through AI2's fully open post-training stack—data, tuning, and code.
| Specification | Details |
|---|---|
| Modality | Text + Image |
| Best For | Visual QA, captioning, image-grounded generation |
| Available Sizes | 7B (32.1GB) |
| License | Apache 2.0 |
This multimodal model bridges the gap between visual and textual understanding, enabling tasks like visual question answering, image captioning, and text generation grounded in images.
| Specification | Details |
|---|---|
| Modality | Image + Text |
| Best For | OCR workflows, document parsing |
| Available Sizes | 7B (16.6GB) |
| License | Apache 2.0 |
Specialized for document processing, this model combines computer vision and natural language understanding to extract and interpret text from images. It's suited for forms, scanned documents, and structured text workflows.
| Specification | Details |
|---|---|
| Modality | Text |
| Best For | Efficient inference (Mixture-of-Experts) |
| Available Sizes | 7B (13.9GB) |
| License | Apache 2.0 |
This model uses a Mixture-of-Experts architecture—an advanced approach where only selected expert subnetworks are activated per input. This significantly improves computational efficiency while maintaining high performance, making OlmoE valuable for real-time or resource-constrained use cases