Vote: what runs first

The machine arrives in late October. Delivery week is one box and a queue, and the front of that queue is yours to decide.

Each option answers a different question, and every one of them is a question we could not find a published answer to for a single 512 GB Mac. The winner is benchmarked first, in full, and its raw JSON is published free. The rest follow in vote order.

Pick one Kimi K2.6, 2-bit 345 GB
Does a trillion parameters at 2-bit beat 400B at 4-bit? Nobody has published this comparison on one machine. Qwen3.5-397B-A17B, 4-bit 224 GB
The comfortable case: a big MoE with room left for 128K of context and a second model resident. GLM-5.2, 4-bit 467 GB
The ceiling. The largest 4-bit build that fits at all, with almost nothing left for the KV cache. DeepSeek-V4-Flash, Q4/Q8 173 GB
The likely daily driver, and a 1M-context claim we can find the real edge of. gpt-oss-120b, MXFP4 63 GB
The speed ceiling, and the Neural Accelerator test: does prefill really double on M5? Llama 4 Maverick, 4-bit 245 GB
The best-documented M3 Ultra baseline, so the generational delta reads clean.

Leave this empty

Add me to the results list (email is not sending yet — see the front page; the feed is the live channel)

One vote per address; voting again changes your pick. We store the address and the choice, and nothing else — not your IP, not a cookie, and not you on the mailing list unless you tick the box. Privacy.

Where it stands

Kimi K2.6, 2-bit0
Qwen3.5-397B-A17B, 4-bit1
GLM-5.2, 4-bit0
DeepSeek-V4-Flash, Q4/Q80
gpt-oss-120b, MXFP40
Llama 4 Maverick, 4-bit0

1 vote, counted live from the database this page reads.

The fallback, stated in advance so it cannot be invented later: fewer than 25 votes by 20 October and the operator picks the order from the predictions post instead. The tally stays public either way, and every model on this ballot gets run.

Why these six and not others: what actually fits in 512 GB. How they will be measured: the method.