Measured answers
The questions people actually ask, phrased the way they ask them. One-sentence answers here; every number, scope and raw run lives one click away. Questions we have not measured say so.
"Enough" splits into three measured experiences: the wait before the first useful word, the flow after it, and whether an answer arrives at all — one round number cannot carry them.
The evidence →The reasoning stream starts instantly but is not addressed to you; the first visible word waits for it to finish. Measure to the first answer token, or the dashboard will call the silence fast.
The evidence →The model reads the whole document before saying anything, and that prefill bill scales with length — measured at two depths on the same box with everything else frozen.
The evidence →The prompt cache: a byte-identical prefix pays the prefill once. Anything that rewrites the prefix — reshuffled RAG chunks, a timestamped system prompt — silently makes every question pay it again.
The evidence →The 35B-MoE class runs with a sub-second first word and room to spare; the frontier 284B class does not fit in one box — that took a pair. Measured, not spec-sheet arithmetic.
The evidence →Not to our gates, on this model and these workloads — while the decode-pace cost of Q8_0 is real and permanent. The uncomfortable finding: the default "run the biggest quant" advice spent speed on nothing detectable.
The evidence →Vulkan decoded faster in our configuration, with correctness identical on every gate — picking the wrong backend cost about as much as doubling the weight precision.
The evidence →We measure it with unanswerable control questions: one model family declined every impossible question, the other sometimes failed by silence — not by fabrication. Both behaviors are invisible unless you test for them.
The evidence →The measured ladder at one, four and eight simultaneous chats shows the wait growing nonlinearly while completion holds — size from the wait your users tolerate, not from a marketing seat count.
The evidence →Three continuous hours under closed-loop load on two identical units: the pace of minute one was the pace of hour three. The predicted burst-then-fade never arrived.
The evidence →A reasoning model can spend its whole token budget thinking and emit nothing addressed to you — transport reports success, the user gets a blank. We measured how often, per model and budget.
The evidence →For chat and structured automation on our box: yes — thinking cost one to two orders of magnitude of waiting and bought nothing the gates could detect. For genuinely hard reasoning, our workloads are not your evidence.
The evidence →Reliability turned out to be a property of the model, not the settings: quant, backend and reasoning mode did not move task success — switching the model family did.
The evidence →Head to head under the same budget and corpus: the model that reasons longer runs out of room before saying anything far more often — and both failures arrive as HTTP 200.
The evidence →Pick the workload first: solo assistant, small team, document-heavy work and the frontier model class are four different purchases. We answer from what our own boxes measurably did — and name the lanes we never measured instead of filling them with arithmetic.
The evidence →On Linux the models live in GTT, not in the BIOS-dedicated slice, and the ceiling is a TTM kernel parameter rather than a firmware menu. Our serving node keeps the dedicated slice tiny and still holds a hundred gigabytes of weights.
The evidence →Different tools the market cross-shops: a measured workhorse for the model class most people run, versus the only desk-format path we measured to the largest open MoE class — as a pair, with seams.
The evidence →It served a 284B MoE at home and walked a context curve to a million tokens; it is also real operational work with half-working stock monitoring. The measured buyer answer names who should and should not buy.
The evidence →dcgm-exporter does not work on GB10 and NVML answers N/A on much of what you alert on; the reliable path is parsing nvidia-smi text through a textfile collector, with a staleness alert on the collector itself.
The evidence →Ask the running server what it runs, pin everything byte-level, freeze the workload, count the right tokens, keep failures in the denominator, repeat, and publish the config next to the number.
The evidence →Because a published number is a config in disguise: unstated workload profiles, defaults that moved between versions, firmware nobody named. Most disagreements dissolve once each figure is pinned to its configuration.
The evidence →Not measured by us yet. Community figures diverge widely with no pinned digests; the qualification run is planned. Until then this page will not quote anyone’s number.
The evidence →A declared lane on our testbed with zero published runs — so zero claims. When the MLX-versus-llama.cpp cell lands, it will land with raw runs.
The evidence →The most asked and least measured question below this hardware class. The offload study is a declared lane; nothing is published, so nothing is claimed.
The evidence →