https://pkillsleep.dev/en/i test things at home (local ai, self-hosting, second-hand hardware), i write down the numbers, and i tell you what works and what crashes.en https://pkillsleep.dev/en/guides/llama-cpp-premier-llm-local/https://pkillsleep.dev/en/guides/llama-cpp-premier-llm-local/Thu, 22 Oct 2026 00:00:00 GMTcompile llama.cpp for your nvidia card and run a recent model with 256K of context on 12gb, while understanding what each option does https://pkillsleep.dev/en/guides/maya-glm-53/https://pkillsleep.dev/en/guides/maya-glm-53/Thu, 22 Oct 2026 00:00:00 GMThow to install project maya, pick the right version of the model when your cards have 24gb or less, and the two options that got me 22% more speed, with what they cost https://pkillsleep.dev/en/guides/strata-deux-gpu/https://pkillsleep.dev/en/guides/strata-deux-gpu/Thu, 22 Oct 2026 00:00:00 GMThow i install strata and run qwen 3.8 flash next on two consumer cards at more than 110 tok/s solo, then how i share it between two agents at the same time https://pkillsleep.dev/en/guides/undervolt-3090-3080ti-linux/https://pkillsleep.dev/en/guides/undervolt-3090-3080ti-linux/Thu, 22 Oct 2026 00:00:00 GMTthe step by step i use on my box: lock the clock, shift the voltage/frequency curve with NVML, check under real load, then make it automatic at boot https://pkillsleep.dev/en/labo/deux-agents-44-66-tok-s/https://pkillsleep.dev/en/labo/deux-agents-44-66-tok-s/Tue, 20 Oct 2026 00:00:00 GMTsame machine, same model, one night of tuning: when two agents generate at the same time, each one goes from 43.8 to 66.4 tok/s, and four settings did most of the work https://pkillsleep.dev/en/labo/glm-53-maya-3090-3080ti/https://pkillsleep.dev/en/labo/glm-53-maya-3090-3080ti/Tue, 20 Oct 2026 00:00:00 GMTtwo maya options give 22% more decode to a 321B model that spills way over into ram, but on a real agent task it stays clearly slower than qwen 3.8 flash next on the same box https://pkillsleep.dev/en/labo/spark-4b-256k-3080ti/https://pkillsleep.dev/en/labo/spark-4b-256k-3080ti/Tue, 20 Oct 2026 00:00:00 GMTspark-x2.6-4b in q8 with llama.cpp: 136 tok/s empty, still 35 tok/s with 254K tokens of context, and a 12gb card that beats the 3090 because i had capped it less https://pkillsleep.dev/en/labo/undervolt-3090-3080ti/https://pkillsleep.dev/en/labo/undervolt-3090-3080ti/Tue, 20 Oct 2026 00:00:00 GMTtwo coding agents working at the same time, 452 W → 358 W across both cards, the fans go from 91% to 48%, and throughput loses 7% https://pkillsleep.dev/en/labo/400k-contexte-80go/https://pkillsleep.dev/en/labo/400k-contexte-80go/Mon, 19 Oct 2026 00:00:00 GMTtwo slots of 409,600 tokens fit on 80gb of ram, and a 373,723 token prompt with three codes hidden inside gets 3 right answers out of 3, read in 195 seconds https://pkillsleep.dev/en/labo/512k-machine-figee/https://pkillsleep.dev/en/labo/512k-machine-figee/Mon, 19 Oct 2026 00:00:00 GMTat first i blamed a dead wifi dongle, but it was the ram: the session cache at 512K needed ~23gb more on top of a 40gb model that was already taking up the space, and a simple memory cap would have avoided it all https://pkillsleep.dev/en/labo/strata-020-ampere/https://pkillsleep.dev/en/labo/strata-020-ampere/Mon, 19 Oct 2026 00:00:00 GMTa simple engine update takes solo decode from 91.7 to 107.6 tok/s on a 3090 + 3080 Ti, and yet the total time of my agents doesn't move, because they spend their life reading prompts