The obvious read on the week's local-inference signal is a hardware story. AMD put a workstation configured for large local models on the price list, someone served a 285-billion-parameter vision model on a rack of consumer 3090s, and a 64-billion-parameter image generator arrived in four-bit builds small enough for a single high-VRAM card. The read that survives contact with the week's own evidence: this was a week about instrumentation, and about how little anyone can currently trust the numbers they are using to justify the hardware.
The finding that reorders the rest came from a developer chasing inconsistent benchmarks on Windows. A local model server attached to a console window held 130 to 200 tokens per second while that terminal had focus and fell to roughly 50 when it did not. The cause sits in Windows foreground scheduling applied to the console host, so neither the model nor the serving engine was at fault. The measurement was taken in the throttled path because the person taking it had clicked over to a browser to run the test, which is what everybody does.
Follow that through and a good deal of published comparison work on Windows is off by a factor that swamps the differences being compared. An engine that looks forty percent faster than its rival may simply have been the one measured in the foreground. None of those results carry a label saying which path they came from. The benchmark wars of the late 1990s ran on the same fault, and I was reading that trade press as it happened: the number that traveled was almost never measured on the configuration anyone actually ran.
The second thread is that the harness now moves capability further than the weights do. A 285-billion-parameter mixture-of-experts vision model became servable on older Ampere cards because someone assembled a compatible serving stack with four-bit experts and eight-bit attention. A million-token context window arrived on Apple silicon through a port rather than a new model. GLM 5.3 Flash posted usable decode and prefill figures on an M3 Ultra at four-bit quantization. Same weights, different plumbing, different machine class. Jensen Huang, talking to Tom's Hardware, framed the economics of that layer plainly: "You sell a chip one time, but when you build software, you maintain it forever." The maintained layer is where the week's gains actually came from.
The third thread runs through what hosted model relationships leave unwritten. DeepSeek appears to have pulled V4 Pro out of active rotation with no deprecation notice, on the same day the Flash-Vision serving recipe surfaced, so anyone pinning that model identifier learns about the change from error logs rather than a changelog. Alongside it, Terence Tao warned that good open problems are being consumed as a non-renewable resource, and a forum thread leveled an unproven charge it called surveillance plagiarism: that hosted providers train on the research sessions users run through them and later present the results as unassisted. Both complaints describe the same missing document. Monday's piece on treating prompts as versioned packages assumed a stable target underneath, and this week the target moved without a version number.
The fourth thread is the quiet argument about whether to buy anything at all. A one-bit 27-billion-parameter model now runs in a browser tab at 25 to 30 tokens per second on a laptop 3060 with 6GB of memory. The interesting part there is the delivery path, since one-bit quantization degrades output badly; the demo buys zero-install access at the cost of fidelity. Desert Ant Labs is pitching on-device inference as a product primitive. Against the new workstation listing, a counter-post argued against letting gear anxiety drive purchases while capable models keep shrinking to fit hardware people already own. The operator question underneath is uncomfortable: how would you know which side of that argument you belong on, when the throughput figure you would use to decide is the one the week showed to be unreliable?
All four sit on the same layer, the software between the silicon and the weights, and that layer is the one nobody is instrumenting carefully. The hardware conversation is legible: memory, bandwidth, dollars. The harness conversation is not, and it moved further over the past few days than any of the hardware did.
Two numbers carry the week: 130 to 200 tokens per second with the window focused, about 50 without. Every other figure in the week's signal, the 285-billion-parameter model on ten 3090s, the million-token context on a Mac, the browser demo running on 6GB, is downstream of whether whoever produced it knew which of those two paths their machine was on.