256GB Mac Studio Beats DGX Spark by ~2x Running Qwen3.8-Flash Agent
GabGarrett · x · 2026-10-21
A hands-on comparison on the same voxel-world agent prompt: a 256GB Mac Studio (oMLX 0.7.0, oQ4e quant with MTP) ran Qwen3.8-Flash-Next in 58.1 agent minutes, writing 225,892 tokens, while a single 128GB DGX Spark (TensorFold, NVFP4 with MTP, MiaAIlab's one-Spark recipe) took 108.6 minutes with 134,322 tokens. Browser check scored 8.5/10 on both; both passed 262,144-token context. Speed benchmark (thinking off): 152 tok/s decode on Mac vs 42.8 tok/s on Spark. Both ran as agents in Qwen Code 0.25.2, thinking on, 131,072-token answer limit; the Mac run was cut short at 58 min by Qwen Code's tool-call cap. The browser check only verifies the page works, not how the world looks.
More from Infra
- Crusoe unveils Tinker API-compatible LLM post-training platform at PyTorch Conference — PyTorch · 2026-10-22
- Open PR optimizes DeepSeek-V4.1 inference on RTX PRO 6000 Blackwell — AlpinDale · 2026-10-22
- vLLM adds day-0 support for Flint: 500 TPS on a single GB200 — _rockt · 2026-10-22
- Grass founder on the Delphi Podcast: training data is now a profitable business serving top labs — grass · 2026-10-22
- DeepSpeed details AutoSP, Muon optimizer, and AutoTP/AutoEP advances at PyTorchCon — PyTorch · 2026-10-22
- Eureka Machine writes a Metal megakernel fusing Flint's entire decode pass — cong_ml · 2026-10-22