models

To choose an LLM model for Tinker, identify your task type from chat, fine-tuning, vision, or research, then match it against a catalog of Qwen, Llama, and Nemotron families and their specific architectures.

Dense and MoE architectures represent different model structures cataloged for Tinker. MoE models use a mixture of experts for efficient processing, while dense models offer a uniform parameter distribution across model families.

You get the recommended renderer for a specific model by invoking the lookup method model_info.get_recommended_renderer_name(model_name), which maps the selected model to its corresponding renderer.

Yes, Tinker supports vision models for fine-tuning tasks. The model catalog includes vision model types with corresponding renderer mappings designed to support chat, fine-tuning, and vision tasks.

Yes, Tinker supports instruction and reasoning model types. The model catalog covers base, instruction, reasoning, hybrid, and vision types across Qwen, Llama, and Nemotron families.

The best way to map a base LLM to a renderer for chat tasks is to use the model_info renderer mapping guidance, which aligns the selected model type with its corresponding renderer for your specific use case.