MiaAI MTP-378.1 tok/s
12.804097 ms/token57.216117% of ceiling
MATCHED INFERENCE EVIDENCE / LOCAL ONLY
Which claim survives the decode gap?
Parse the supplied recipe summary, reproduce the timing bridge, and keep measurement, mechanism, and missing evidence in separate lanes.
SOURCE SUMMARY
Two rows. One matched context.
No benchmark serviceValues are source-supplied or editable. No GPU, model run, live profiler, account, wallet, or social action.
CLAIM LEDGER EMPTY
A faster row is not yet a causal explanation.
Audit matched context before comparing throughput. The completed state will show exact timing, bandwidth context, an implied read model, and evidence that remains unavailable.
VERIFIED COMPARISON
MATCHED DGX SPARK SUMMARY
MTP-3 leads this supplied run.
speed gap+7.3 tok/s
relative gain+10.310734%
time saved / 1K1.320196 s
bandwidth-only ceiling136.50 tok/s
0273 GB/s / 2 GB per token136.50 tok/s
Unsloth NVFP4-Fast70.8 tok/s
14.124294 ms/token51.868132% of ceiling
equal-overhead inference0.906530 relative full reads9.346991% implied reduction
78.1 / 70.8 = 1.103107
Same 4-bit width; MTP depth differs; no ablation.
7.6 vs 7.3 · 5 stories · paired scores absent
Runtime loading. No benchmark audited.
match contextparse rowscompute timingmodel bandwidthlabel inferencewithhold missing statistics
TRANSFER THE METHOD
A claim can be numeric and still outrun its evidence.
The two supplied rates support an exact recipe-level speed and latency comparison for this stated run.
Relative reads translate the gap into the post's bandwidth story only under equal non-weight overhead.
Paired significance needs per-story differences or sufficient dispersion; two means and n=5 do not provide them.