On RTX 5090 and DGX Spark, Foundry Local’s agentic coding—where AI autonomously breaks down tasks, calls tools, and modifies code—ran 35% to 135% faster in testing. Our judgment is that AI coding is moving from “suggestions in a chat window” toward “executing tasks on your own machine,” but it remains a considerable distance from affordable, accessible use for everyone.
What this is
Foundry Local is Microsoft’s solution for running models on local devices. This discussion focuses on its agentic coding capabilities: instead of merely generating code snippets, the model can perform multistep operations such as analyzing a project, calling tools, modifying files, and checking results. The tests ran on RTX 5090 and DGX Spark; the community subsequently asked whether these capabilities could be added to the open-source llama.cpp tool, a local inference framework for running large language models with relatively low resource usage.
Industry view
The value of local execution goes beyond speed. Code and business data can remain on the device, reducing the privacy concerns associated with cloud uploads and allowing work to continue when the internet is unavailable or cloud services are congested. For enterprises, deploying models in a controlled environment may also lower the cost per task.
The risks are just as clear: the 35% to 135% improvement comes from tests on specific hardware and under specific conditions, so it cannot be extrapolated directly to ordinary computers. High-end GPUs and workstations remain a barrier. Agentic coding also introduces issues involving accidental file changes, incorrect execution, and permission management. The community’s request to port these capabilities to llama.cpp shows that an open ecosystem remains a key variable. Whether Microsoft can make the toolchain, stability, and cross-device experience genuinely work smoothly still needs to be proven.
Impact on regular people
For enterprise IT: Processing sensitive code locally is attractive, but enterprises must also assess compute procurement, permission isolation, auditing, and deployment and maintenance costs. Pilots should start with low-risk tasks.
For individual professionals: Local agents can reduce waiting times and may keep materials off external services. However, people without high-performance hardware are still unlikely to get an equivalent experience in the near term.
For the consumer market: AI coding will gradually enter desktop software and developer tools, but “it runs on my machine” does not mean “an ordinary user can use it.” Price, ease of use, and security mechanisms will determine how quickly it becomes mainstream.