Breakage
Amber means it bit. These are measurements and one documented mess, not fixes I invented afterward.
warn
With GPT-OSS loaded, September 2026 testing left about 708 to 717 MiB of VRAM on the RTX 5060 Ti.
MeasuredWhat it taught meThere is not a spare pile of memory sitting next to that model. A second big load is how this card falls over.
warn
GPT-OSS passed 9 of 16 current-contract tool tasks in one controlled evaluation.
MeasuredWhat it taught meThe tokens-per-second number was not the job. Nine out of sixteen is what that tool contract actually scored.
warn
A prompt of about 43K tokens took 14.6 seconds of cold prefill, and about 21 milliseconds once the prefix was cached.
MeasuredWhat it taught meThe warm number is the cached one. It is not how long the prompt takes the first time.
warn
In the ~120K-token test, GPT-OSS generated at 61.3 tok/s, against a ~95 tok/s baseline.
MeasuredWhat it taught meThe model took the long prompt. It did not keep the short-prompt speed.
warn
One WSL/Docker fix wiped the running containers and forced a major rebuild.
DocumentedWhat it taught meThe repair that got the box booting also deleted the work. Story 4: The pain