# Microsoft evaluates different open models for Cowork > Source: > Published: 2026-06-18 14:40:00+00:00 TestingCatalog has learned that **Microsoft teams behind Copilot Cowork are evaluating a broader set of open and open-weight models beyond DeepSeek** as potential underlying options for the agentic work app. [Axios recently reported](https://www.axios.com/2026/06/16/microsoft-copilot-cowork-tokenmaxxing-cowork?ref=testingcatalog.com) that Microsoft is considering a Microsoft-hosted version of DeepSeek as a cheaper model option for Copilot Cowork, but the evaluation appears to extend beyond a single model family. The key architectural detail is the separation between the model layer and the harness. Copilot Cowork is being built so the orchestration system can remain stable while the underlying models are swapped depending on task type, cost, latency, and required capability. That would allow Microsoft to route some work to frontier APIs, while other parts could be handled by cheaper self-hosted models on Azure. Over time, some lighter tasks may also become candidates for local execution as smaller models mature. This direction would make sense for [Copilot Cowork](https://www.testingcatalog.com/exclusive-new-screenshots-of-upcoming-copilot-super-app/), which is now [generally available](https://www.testingcatalog.com/microsoft-launches-copilot-cowork-globally-for-microsoft-365/) for Microsoft 365 Copilot customers and is aimed at enterprise users who expect reliability, governance, and data controls. For customers, the main benefit would be more flexible pricing and model choice. For Microsoft, it offers a way to reduce dependence on any single external model provider while controlling compute costs for long-running agentic workflows. At the same time, this creates an internal competition. Microsoft has been presenting its own [MAI models](https://www.testingcatalog.com/microsoft-build-2026-recap-from-windows-to-copilot-all-ai/) as part of a broader push to become a top-tier AI lab next to OpenAI, Anthropic, and Google. If Chinese and open-weight models continue to perform strongly enough to power parts of Copilot Cowork, Microsoft’s internal model teams will face a clearer benchmark: they need to compete with fast-moving open-model providers. **According to sources familiar with the evaluation, these open-model developments have not reached production in Copilot Cowork yet**. For now, they remain under active testing. The practical question is not whether Microsoft can plug in another model, but which model can meet enterprise expectations once cost, compliance, safety, and task quality are measured together.