Foundations Model, checkpoint, artifact, runtime A precise vocabulary for separating model identity from the files, software, and hardware used to run it. 8 min →
Architecture Dense models and mixture of experts How expert routing changes the relationship between total parameters, active parameters, memory, and compute. 9 min →
Weights The post-training pipeline How continued pretraining, supervised fine-tuning, preference optimization, and reward-driven training turn a base model into a usable release. 13 min → Fine-tuning, LoRA, and QLoRA How full-parameter and parameter-efficient tuning update a model, what an adapter contains, and how training choices affect deployment. 14 min → Knowledge distillation What a student model learns from a teacher, what is actually transferred, and why a distilled model is a new checkpoint. 8 min → Quantization without the shorthand How lower-precision representations use scales and groups, why 4-bit formats differ, and what quality claims require. 11 min → NVFP4 and block-scaled 4-bit inference What NVFP4 stores, where the extra half-bit estimate comes from, and why Blackwell support is more than VRAM capacity. 9 min →
Deployment Model memory, KV cache, and context A practical memory model that separates weight payload from cache, temporary buffers, concurrency, and runtime headroom. 10 min → How to read a model repository A file-by-file method for checking lineage, completeness, configuration, weights, tokenizer, license, and evaluation evidence. 12 min → Choosing a local inference runtime How to choose between desktop applications, portable engines, and GPU serving stacks without treating them as interchangeable wrappers. 10 min →