How do we run large models on accessible hardware while preserving their ability to do useful work? Our living white paper records the design, experiments, failures and unresolved questions behind Krasis.
Living white paper · Version 0.5 ·
Phase-Separated Inference for Large Mixture-of-Experts Models on Memory-Constrained NVIDIA GPUs
Full GPU prefill, measured expert residency and native Rust/CUDA decode, alongside Agentic Development Qualification (ADQ), quantisation findings and conversation-cache correctness.
Research status is explicit: the archived GLM-5.3-Flash campaign passed 2/4 development tasks, and the current conversation-loading repair remains unqualified.
AI-assisted drafting from engineering records and experimental data, with human review and curation. Northloom Ltd takes responsibility for the published material. Version history is recorded in the paper's appendix.
