What to run on 256 GB Mac

2 complete runs

#	Creator	Setup	Overallof 100	Agent testbrowser check	Speedtok/s	Qualitymean of 7 tests	Memorylargest context
1	oMLX (Jun Kim)	Qwen3.8-Flash-NextnewMac Studio M5 Ultra	87.3 of 100rank 1 of 2as of 20 Oct 2026	8.5 / 10	152.5≈ 5.4 s for 500 wordscode 175.8 · prose 129.1	64.8%mean of 7 tests	256Ktokens tested
2	gnahzhang	DeepSeek-V4.1-Flash-1017new1× Mac Studio M5 Ultra	64.6 of 100rank 2 of 2as of 20 Oct 2026	8 / 10	63.8≈ 11.2 s for 500 wordscode 67.7 · prose 60.0	60.0%mean of 7 tests	128Ktokens tested

Creator: who made the recipe; the picture leads to their profile. Overall: 0 to 100 from speed (45 %), quality (30 %), the browser check (5 %) and my rating (20 %); the run with the highest score is rank 1 and my pick. The score shows the date of the newest measurement in the class. Scores are relative to the best run in the class today; a better model later moves older runs down. How the score is made. Agent test: the model builds one 3D scene as an agent, one attempt, no help; a browser checks the page, 0 to 10 points. Speed: tokens per second with one request, thinking off. Quality: the mean of 7 tests (MMLU-Pro, GPQA, IFEval, LiveCodeBench, long text 8K and 32K, GPQA with thinking), each with its 95 % margin of error on the recipe page. Memory: the largest context that loaded and answered. All other numbers (each quality test, time, peak memory, sparkDash, engine, quant) are on the recipe page. Only a run with every part done and independently checked is listed. Click a row for the recipe, the exact command and the raw files.