Justine Tunney created Cosmopolitan Libc and llamafile, open-source projects that package sophisticated software—including large language models—as portable executables. By combining cross-platform distribution with faster CPU inference, she makes advanced models practical on ordinary computers without complicated installation or expensive dedicated hardware.
From Google to portable computing
At Google, Tunney contributed to the domain-registry platform Nomulus, the Bazel build system, TensorFlow, and TensorBoard, where she served as a technical lead. In 2016, she organized Operation Rosehub, using BigQuery to identify approximately 2,600 GitHub repositories dependent on a vulnerable Java library and coordinating roughly 50 Google employees to help submit fixes.
After leaving Google in 2018, Tunney concentrated on independent open-source development. Her Cosmopolitan C library produces native programs that run across Windows, macOS, Linux, and multiple BSD systems. Its Actually Portable Executable format combines otherwise incompatible executable conventions inside a single binary. She applied the same architecture to Redbean, a self-contained web server capable of carrying its application and assets inside the executable.
In November 2023, Tunney and Mozilla introduced llamafile, combining Cosmopolitan with llama.cpp to distribute model weights and inference software together. The open-source project lets users run compatible models locally across supported operating systems while retaining control over their data and operating without an internet connection.
The ideas driving her work
Single-file local AI: Tunney builds executable files that bridge operating-system conventions, including a Unix shell entry point embedded in a Windows portable executable. One artifact can contain both a model and its runtime, eliminating platform-specific installation and dependency assembly.
Faster CPU inference through matrix multiplication: Tunney developed specialized matrix-multiplication kernels that accelerate prompt processing, including long-context tasks such as summarization. Her technical account of the optimization work describes 84 kernels spanning processor families and techniques including outer-loop unrolling, improved cache behavior, and GPU-inspired synchronization on CPUs.
Open acceleration with tinyBLAS: Her tinyBLAS implementation supports GPU execution without requiring applications to bundle heavyweight proprietary numerical libraries or depend on a separately installed development toolkit. Tunney treats distributability and openness as engineering requirements alongside raw speed.
Larger models on accessible hardware: Consumer CPUs can use relatively inexpensive system memory to accommodate models that exceed the dedicated memory available on many graphics cards. Her AI Engineer World’s Fair presentation demonstrated how this tradeoff can expand access to capable local models even when generation takes longer.
Authorship and open-source provenance: In her writing about AI training and software history, Tunney argues that training pipelines should preserve developers’ names, copyright notices, and version-control history. She regards attribution as essential to sustaining independent programmers whose publicly shared work becomes training material.
