Descriptions:
Bart Slodyczka runs DeepSeek V4 Flash 0731 on a Mac Studio M3 Ultra with 512GB unified memory, testing whether a locally-hosted open model can handle real agentic workflows. According to the Frontier Language Model Intelligence Index — covering nine evaluations including agentic coding and knowledge work — DeepSeek V4 Flash scores 52, on par with GPT-5.4 from March 2026 and ahead of Claude Sonnet 4.6 at 48. Slodyczka is candid that he is usually disappointed when benchmark claims meet reality, but reports genuine surprise at this model’s performance.
Running the 4-bit MLX quantized version (167GB weights) inside the pi.dev agent harness with VS Code, he achieves 41.7 tokens per second with multi-token prediction enabled versus 26.1 t/s without — a meaningful real-world speed difference. He notes the 1-million-token context window may be memory-constrained on 256GB devices and points out that two NVIDIA DGX Spark units (~$10K USD total) can run the model at a reported 72 tokens per second for those who can’t source Apple’s high-memory hardware.
The three live tests — fixing a broken n8n workflow with Brave web search authentication, completing a vague reporting task from an Excel sheet with stock and sales data, and a ClickUp automation — show DeepSeek diagnosing multi-step problems autonomously, installing required dependencies unprompted, and producing working outputs with minimal instruction. The video is a practical capability demonstration rather than a promotional overview.
📺 Source: Bart Slodyczka · Published August 10, 2026
🏷️ Format: Hands On Build
