AI measured, mostly local LLM inference on hardware I own. Every run comes with its methodology and run log, numbers you can reproduce.
The same discipline on every run: fixed seeds, temperature 0 where it matters, first result kept, every number traced to a log line or a file on disk, and n disclosed wherever a number is an estimate. Most runs are local, one RTX 5090 and one vLLM server. One is a closed access image model, held to the same rules. Four runs are posted so far, each asking one question.
Does an image editor remember the image? Krea 3, GAUNTLET Edit takes one photoreal frame through twelve chained edits (night, snow, spring, a storm, a woodblock print) and then asks for the original back. Every edit landed first try and a sign stayed legible through all of them, but the return moved 0.3% of the pixels and stayed 40% off. It is on video, with the drift scored per step.
Does thinking mode earn its tokens? Gemma 4 31B, thinking off vs on gives a 31B model one prompt, one seed and two tries at building a procedural 3D island in Three.js, recorded. Thinking cost 3.8 times the tokens and 93 extra seconds, and turned a pile of shards into an island with 126 trees. The write-up is about when that trade is worth making.
How much can one card actually hold? The 5090 receipt run feeds a 27B model all of Frankenstein, 103,459 tokens, in a single prompt. It took four undocumented fixes to get there, and the page is written as a receipt: every figure sourced to the vLLM log, plus lm-eval probes (gsm8k, wikitext) so the quantised model is checked for quality, not just speed.
Can a 32B model get faster without a second model? Qwen3-32B, three configs runs the same 24 prompts across a baseline, n-gram speculative decoding and FP8 KV cache. One flag took code edits from 75 to 350 tok/s. Determinism was checked rather than assumed: two identical baseline runs had to match byte for byte before a number counted, and that check is why FP8 KV cache stays off on my server despite its throughput win.
Each run page has the full methodology, every result, and the log lines the numbers came from.
One photoreal frame, twelve chained edits, and one instruction to put it all back. Krea 3 landed every additive edit first try and kept a sign legible through a woodblock print, then came home with a different woman on a different mountain and could not undo a thing. Scored by pixel drift, on video.
One prompt, one seed, two runs of a local Gemma 4 31B: build a procedural 3D island in Three.js from scratch. Thinking cost 3.8 times the tokens and 93 extra seconds, and bought a real island instead of a pile of shards. What reasoning fixes, what it cannot, and when to switch it on. Both ran first try. n=1, recorded.
I fed a 27B model all of Frankenstein in a single prompt on one RTX 5090. It took four undocumented fixes, and every number on this page is sourced to a log file on disk.
A 32B model on one RTX 5090 went from 75 to 350 tok/s on code edits with a single flag and no second model: n-gram speculative decoding. Same 24 prompts across three server configs, with determinism checked byte for byte rather than assumed.