Efficient inference
Low-overhead runtimes, careful memory lifecycle, and reproducible latency and throughput measurements.
Research-grade tooling for local inference
We build open-source software alongside research in language models, diffusion, and systems ML—with an emphasis on measurable performance, honest benchmarks, and sustainable engineering.
What we work on
Focus areas
The Foundry brings together systems-minded contributors who care about what happens on real hardware—not only on slides.
Language & generative models
Tools and experiments grounded in current LLM and diffusion research, shared as open code and docs.
Open collaboration
Contributions welcome where evidence and scope are clear—see each repository for contribution guidelines.
Repositories
Projects
Active and upcoming initiatives under the Inference Foundry GitHub organization.
super-ollama
Terminal-native, in-process local LLM engine (no HTTP in the main UX); llama.cpp via CGo, focus on low overhead and clean teardown.
Crucible
Open research journal and experimental log.
BitForge
Quantization theory, methods, and reproducible experiments across bit-widths and runtimes.
Lexicon
Open fine-tuned prompt catalog with versioning, licensing, and analysis for reuse.
Argus
Algorithms to detect AI-generated images using JEPA-based representations.
How we work
Principles
These expectations apply across our public spaces and reviews.
- Evidence over opinion. Performance claims belong with methodology: hardware, versions, and raw artifacts when possible.
- Safety and respect. Follow the organization Code of Conduct; debate the technical content, not the person.
- Scope discipline. Large features start with discussion in the relevant repo so maintainers and contributors share context early.
Contact and collaborate
For contributors and collaborators, use Discord first. For direct contact, message the founders through GitHub or LinkedIn.
Join Discord