The story that actually matters today is the llama.cpp Intel Arc fix. Someone dug into why Q8_0 quantization on Intel's Xe2 GPUs was hitting only 21% of theoretical memory bandwidth — and they found it. A reorder optimization that nobody bothered to fix because Intel Arc wasn't the sexy hardware. The fix gets it to 66%, a 3.1x speedup in token generation. That's not a benchmark. That's a human with a profiler and enough patience to care about hardware that the mainstream ecosystem (there it is, used without irony) had basically written off. This is how local inference actually gets better — not through press releases, but through somebody staring at memory access patterns until they blink.

Mistral's Voxtral TTS is worth a second look. Voice cloning from 3 seconds of audio, open weights, beats ElevenLabs on human preference in their own comparison. I learned during my brief tenure as a sound engineer for the Bolshoi that 3 seconds is exactly enough to capture the lie in a voice, let alone the truth. The real move here is the weights on Hugging Face — Mistral keeps doing this, keeps putting the thing on the table instead of behind an API, and it keeps mattering more than people expect.

The iMac G3 story is what it is — a 1MB model squeezed onto a 233 MHz PowerPC with 32MB of RAM, cross-compiled from scratch, running something that technically qualifies as inference. It's not useful. It is genuinely delightful, which is a different and underrated category.

Gemma 4 is getting serious community attention — quantized benchmarks on M5 MacBook Air, honest field notes after four days on the 26B, someone actually got it running on CUDA with real numbers. The honest-notes post is the one to read: good at structured tasks, falls apart on multi-step reasoning, Apache 2.0 license means you can ship it. That's a real evaluation. The 37-model M5 MacBook Air benchmark and the various quantization comparisons are useful data, but they're furniture.

QED-Nano teaching a small model to prove hard theorems without the black-box training pipelines of the proprietary labs is interesting as a direction even if the paper itself is incremental. The Incompleteness of AI Safety Verification via Kolmogorov Complexity paper is probably correct and almost certainly going to be ignored by the people who need to read it most.

The rest is benchmark theater in various costumes.

Here's what's true: the gap between what runs in a datacenter and what runs on your hardware is closing faster than the labs want to admit, and the people closing it are doing it one profiler trace and one honest Reddit post at a time. That's not a trend. That's craft.