https://modelfoundry.example/blog.html Engineering notes on quantization, serving, evaluation and supply-chain verification for open-weight models. en https://modelfoundry.example/post.html?p=quantizing-models-without-losing-the-edge-cases mf-note-quantization-edge-cases Perplexity moves by a fraction of a point after quantization. Production incidents do not. A practical checklist for validating Q4 builds. https://modelfoundry.example/post.html?p=serving-70b-moe-on-a-single-workstation mf-note-serving-moe-workstation Expert offload, page cache pressure and the routing behaviour that surprised us on a 24 GB accelerator. https://modelfoundry.example/post.html?p=why-we-publish-an-eval-harness-with-every-release mf-note-eval-harness Benchmark tables age badly. Shipping the harness means a result can be reproduced instead of trusted. https://modelfoundry.example/post.html?p=token-economics-batching-and-the-p99-cliff mf-note-token-economics How continuous batching, prefix caching and admission control interact under a burst. https://modelfoundry.example/post.html?p=fine-tuning-small-models-on-a-budget mf-note-fine-tuning-budget A LoRA workflow that fits on two consumer GPUs and produces adapters worth shipping. https://modelfoundry.example/post.html?p=how-uploaded-weights-get-verified mf-note-weight-verification Hashing, format inspection, tokenizer consistency, smoke inference and the publication gate.