I take apart everything before I throw it out, which is how I learned to fix half of it. Plastic clips, hidden screws, and the correct order to reassemble a thing.
On question 3: stop asking the model. Self identification has been unreliable for as long as there have been point releases, and it is worse right now because the tier names are new enough that the model has barely seen them written down. It will confidently give you a name that is wrong in an interesting way.
Use the label on the message and the regenerate menu instead. Those come from the client, not from the weights. And if you need it reproducible rather than merely correct, that is what the API is for: you pass a model id, you get that model, nothing routes underneath you. The chat product is explicitly trying to be smart about which tier you need. That is a feature if you are chatting and a problem if you are measuring.
Incident hygiene, since you said two person team and no ops: do not roll back a deploy that is not implicated. You have not deployed since Tuesday and this started today, so the deploy is not the cause and rolling it back only adds a variable.
Keep a running note of what you checked and when. Not for the postmortem, for you in forty minutes, when you cannot remember whether you already tested from the hotspot.
One trap when you do this comparison: on eligible paid plans the Instant setting can step up on its own for requests that look demanding. So your "Instant" arm is not guaranteed to be Instant throughout, particularly on exactly the hard prompts where you are trying to show it falls over.
If you want a clean comparison, use prompts that are unambiguously hard and unambiguously easy rather than the interesting middle, or accept that the middle of your results is measuring the router rather than the model.
Worth doing this arithmetic any time a load feels slow. If your measured time is well above file size divided by drive speed, the disk is not the bottleneck and you are about to buy the wrong hardware.
Watch out for the related one: if you serve two models and the second gets requested, the first can be evicted to make room and now you have two reload cliffs alternating. OLLAMA_MAX_LOADED_MODELS controls how many stay resident, but only if the VRAM is actually there for both.
Second the order of operations. I fixed my chair and footrest first and discovered the desk height was fine all along, the problem had been that my feet were unsupported so I was bracing with my legs and locking my shoulders.
Pull the MOT history online from the reg before you even drive out to see it. It is free, it takes a minute, and it tells you more than the seller will. What you are reading for is not the passes, it is the advisories and the mileage line. Advisories that repeat year after year mean nothing has ever been fixed, mileage that jumps around means the recorded readings are inconsistent, and a car with a gap in its testing history was sitting somewhere for a reason.
Second the Q5/Q6 band. The interesting jump is between Q4 and Q5, not between Q6 and Q8 - below Q4 things get noticeably worse and above Q6 you are paying memory for almost nothing.
If the schema is stable and you are doing hundreds a night forever, the actual answer in six months is a fine-tuned 7B. A LoRA on 500 hand-corrected examples will beat both of these on your specific documents and run at 90 tok/s. Not a today decision, but stop thinking of the general model as the destination.
Six weeks, internal, 60 staff who are all on the same corporate laptop image: use whichever you type faster. This decision is worth about half a day of your six weeks and you have already spent more than that thinking about it.
It is a hot pot problem and terracotta is making it worse, since it is porous and evaporates through the walls as well as the surface. Three practical changes: move it so the pot itself is shaded in the afternoon even if the leaves are in sun, mulch the surface with 2cm of compost or bark, and stand the pot on a saucer with a little water in it during heatwaves only. Also, peat-free multipurpose dries much faster than the old stuff and rewets badly once it has fully dried, so if the surface is pulling away from the pot edge your morning water is running down the gap and straight out.
I'll push back gently on where this thread is heading, because 'posture doesn't matter, just move' gets repeated until people stop setting anything up at all. A laptop flat on a coffee table for eight hours is not rescued by standing up every half hour, and neither is a monitor 30 cm to the left of your keyboard. The honest version is: get a sane baseline so none of your positions are actively bad, then vary between them. Both halves, not one.
One thing to watch for: if they set flowers while stalled, pinch them off. A stressed plant that fruits early stays small all season. Let it build leaves first.
I panic bought with three weeks left on an inspection and it cost me. Drove it twice around a housing estate, liked the way it pulled, handed over the money on a Sunday. Two months later the sills turned out to be filled and painted, which an inspection would have caught for the price of a nice dinner. I sold it a year later for a lot less than I paid and I have never again bought a car without putting it on a lift first.
Since the real complaint is swapping: set the container memory limits, and if you are on Linux with fast storage consider zram over a swapfile. Any inference workload that touches swap during generation goes from tolerable to unusable within one token.
The fan does more than prevent damping off, it also thickens the stems. Seedlings that get a bit of air movement or a daily brush with your hand come out stocky and transplant far better than protected ones.
Practical things nobody mentions about a low rear drive car left outside: the boot lid seal freezes shut, the doors freeze if you wash it and do not dry the seals, and heated seats matter more than heated grips ever did on a bike because the cabin takes a while to warm. Silicone grease on the door seals in November is a two minute job that saves you a lot of swearing in January.
Watering daily may actually be part of it. Light daily watering keeps the surface wet and gives roots no reason to go down. Water deeply every third day instead, once they are in the ground, and they root outward much faster.