Let the squirrel brain go to work and create some fun stuff.
How AI Pitlane works: pipelines as validated DAGs where the model behind any node is a config field, budgets that refuse paid calls instead of logging them, a work-order loop that stops at a draft PR — and the 124-tool MCP server that lets an AI client build, launch and monitor the graphs itself.
Google's MTP drafter takes Gemma-4-12B from 78 to 133 tokens per second on one RTX 3090 once it is paired by quantization and vLLM is told to use piecewise graphs. Then the settings: reasoning_effort is a no-op in Gemma 4's template, thinking on runs 4,000 to 8,000 tokens and loops at low temperature, a per-request thinking budget vLLM implements and Google does not document bounds it, and across 41 prompts under four recipes greedy decoding is flawless on 40 and wrong the same way every time on one.
Eight quantized models on one RTX 3090, each given a versioned spec, a thirteen-probe suite with a whole-file edit task, a recipe chosen by measurement and expectations frozen with spread-derived tolerances. The first sweep found the function task bimodal at temperature 0.2 while the edit task held at 7/7, the dense models byte-deterministic at temperature 0 and the MoE model not, and reasoning_effort honoured by one chat template in eight.
A local model's context window is set by VRAM, not by the model card. An agent pipeline that reserved 16,000 output tokens before counting its input, the token budget and per-endpoint model profiles that replaced it, the pre-commit audit that caught a floorless 'fit' turning a loud failure into silent partial commits, and the three faults only the live server could reveal.
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.
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.
No research page matches that search.