LabModern RPG Solutions
Measured tests of local AI
Language models, image and video generation, and the memory engine behind our assistant, run on our own machine with two graphics cards. Each article states its question, its setup, its results and what it does not prove.
Articles
- Basalt on one RTX 5090: IQ3_S writes code 22 to 25% slower than IQ3_XXS with nothing left on the SSDOn one RTX 5090 at 460 W, Basalt's IQ3_S pack of Qwen3.8-Flash-Next wrote code 22 to 25% slower than IQ3_XXS with nothing left on the SSD, 44 to 47% slower with 11% of its experts there, and half as fast with a quarter there. Prompts of 6,756 to 8,576 tokens; three runs, four draws per pack.
- An RTX 5090 at 460 W instead of 600 W: prompts take 9 to 16% longer to read, replies are written 3 to 8% slowerWe had capped an RTX 5090 at 460 W to keep a home server cool and quiet, then measured what that costs a 27B model against the card's 600 W maximum. Reading a prompt took 9 to 16% longer. Writing was 3 to 4% slower without multi-token prediction, and about 8% slower with it for the same share of drafted tokens kept. In August, on another engine without drafting, the same cap had not slowed writing.