Posts
5 posts-
Jun 29, 2026
Model analysis summary
A deep dive into the different parts of the model ids - parameters, quantization, but also architectural stems like MXFP8, NVFP4, MLX, etc. - what they mean and how they interact with each other. Semi-raw analysis from claude code.
ai llm quantization ai-generated -
Jun 29, 2026
Model-ID stem analysis - quality / speed / size rankings and combinability
Analysis of the architectural and quantization stems in model identifiers, ranking them on quality, speed (TTFT and TPS), and size across Mac, Linux, and WSL2 - plus a combination matrix of what stacks and what is mutually exclusive.
ai llm quantization ai-generated -
Jun 29, 2026
Model glossary run 05 - Performance & combination research notes
Run 05 of the model-ID glossary analysis - verified findings and source links on the GGUF quant ladder, MLX vs llama.cpp, NVFP4/MXFP4/FP8, QAT, TTFT vs TPS, and orthogonal build-time techniques.
ai llm quantization ai-generated -
Jun 29, 2026
Model glossary run 06 - Stem axis classification and combinability
Run 06 of the model-ID glossary analysis - classifies stems into the three orthogonal axes (numeric encoding, runtime, build-time technique) and tabulates which combinations stack, warn, or are mutually exclusive.
ai llm quantization ai-generated -
Jun 29, 2026
Model glossary run 07 - v2 verification
Run 07 of the model-ID glossary analysis - re-researches the four hedged claims in the analysis report (Ollama MLX backend, FP4 in llama.cpp, Apple Silicon FP4/FP8 hardware, MLX long-context decode) with sources.
ai llm quantization ai-generated