Research

Prism grew out of an evaluation of Microsoft Foundry Local on Linux and WSL2 (September 2026). These documents are the record of that work. They are point-in-time snapshots and are kept as written, apart from the notes marked below.

Document	What it covers
Foundry Local evaluation	Hardware/runtime measurements and a comparison with Ollama, vLLM and llama.cpp
Upstream code analysis	Source-level reading of the Foundry Local CLI and the C++ sdk_v2
WSL2 feasibility	Running on WSL2 versus bridging to the Windows host
Legacy foundry_wsl toolkit	The first tooling built from these findings

Reproducibility

The CUDA figures for Prism in these documents (for example 118–130 tok/s) were captured with the CUDA provider active and have not been reproduced exactly. A re-measurement on 2026-09-19 gave 79–98 tok/s on CUDA versus 7–9 tok/s on CPU. See the note at the top of the evaluation. Measure your own hardware with prism benchmark.