Working notes on local LLM infrastructure, agent pipelines, and shop hardware — measured, with the data.https://squirrelglitch.com/en-ushttps://squirrelglitch.com/research/3090-bake-off/https://squirrelglitch.com/research/3090-bake-off/Qwen3.8-27B, Ornith-1.5-35B-A3B and DiffusionGemma-26B-A4B, each booted through llama-swap on the same workstation, timed with the same harness, and judged by whether the code they returned actually runs. With a small-model round and a controlled prompt-and-sampling re-test.Sun, 30 Aug 2026 12:00:00 GMTlocal-llmvllmllama-swapbenchmarkhttps://squirrelglitch.com/research/3090-local-coding-stack/https://squirrelglitch.com/research/3090-local-coding-stack/What is actually worth running on a dedicated 24 GB Ampere card in late August 2026, measured rather than advertised: llama.cpp + llama-swap against vLLM W4A16, the Ampere-specific traps, and how it wires into the tools already running on the workstation.Sun, 30 Aug 2026 12:00:00 GMTlocal-llmllama.cppvllmampere