The video introduces the Foundry toolkit within Visual Studio Code, which simplifies AI model selection by providing access to a diverse multi-model ecosystem and integrating GitHub Copilot for tailored recommendations based on specific criteria. It demonstrates a streamlined workflow of shortlisting models, exploring detailed catalogs, and conducting side-by-side testing in a playground environment, enabling developers to make informed, scenario-specific choices without guesswork.
The video addresses a common challenge in building AI applications: selecting the appropriate AI model from a vast array of options across different providers, sizes, and capabilities. To simplify this process, it introduces the Foundry toolkit integrated within Visual Studio Code, which offers a flexible multi-model ecosystem. This toolkit provides access to a wide range of generative AI models, including small and large language models from sources like Microsoft Foundry, OpenAI, Anthropic, Google, GitHub-hosted models, and even local models running via ONNX or Ollama. Users can also integrate their own custom models, enabling them to avoid being locked into a single provider.
To begin model selection, the video demonstrates how GitHub Copilot can be used to generate recommendations based on specific criteria such as deployment region, subscription, and required capabilities like image input support. For example, by specifying an Azure subscription and region, Copilot queries available models that meet these requirements and suggests options like GPT-5 and Claude Sonnet 4.6. It also provides a comparative analysis table and alternative model suggestions, helping users quickly narrow down their choices for scenarios such as marketing with image processing needs.
Next, the video explores the Foundry Toolkit’s model catalog, a centralized hub for discovering and filtering models by various attributes including hosting provider, publisher, features, and deployment options. Users can validate Copilot’s recommendations or manually search for models like GPT-5 and GPT-4o from OpenAI, filtering for image attachment capabilities. Each model card offers detailed information such as key capabilities, use cases, pricing, and technical specifications. The video then shows how to deploy selected models directly within the Foundry Toolkit UI, customizing deployment names and token limits before deploying them to Microsoft Foundry.
Once deployed, the models appear in the Foundry Toolkit’s resource panel, where users can open them in a playground environment to test and compare their performance side-by-side. The video illustrates this by running identical prompts through GPT-5 and GPT-4o, noting differences in response speed, style, verbosity, and handling of image inputs. For instance, GPT-4o responds faster and includes more emojis and hashtags, making it appear more marketing-ready, while GPT-5 offers a more structured and fluent response. The playground also supports attaching images to test models’ vision capabilities, with both models accurately extracting text but differing in formatting and token usage.
In conclusion, the video highlights a streamlined workflow for AI model selection using the Foundry toolkit: starting with GitHub Copilot to generate an initial shortlist, refining choices through the model catalog with filters, and conducting detailed side-by-side testing in the playground. This approach empowers developers to make informed decisions tailored to their specific scenarios without guesswork. The video also hints at a follow-up session that will delve deeper into the agent development process, building on the foundation established here.