Summary
In this Frontier AI at Home episode, sentdex evaluates DeepSeek-V4-Lite 1017, a continued-training release from October 17th, as a local alternative to frontier hosted models. After finding the earlier V4-Lite wasteful with tokens, he reports that the 1017 variant keeps its heavy, iterative thinking but now works until the job is actually finished, which makes it well suited to handing off long tasks.
The video walks through benchmark results. On Terminal Bench V2.1, DeepSeek V4 Flash 1017 scores 92 percent, answering 82 of 89 questions correctly with the OMP harness, which means the benchmark is close to saturated. A harder Terminal Bench 4 test follows, where GLM 5.3 Flash surprisingly comes out ahead, a result that conflicts with published figures and which sentdex attributes in part to the harness. He also notes DeepSeek uses roughly three to fifteen times more tokens per task.
The second half turns to hardware. Sentdex discusses running models on Dell workstations, including a T4 with dual RTX Pro 6000 Max-Q GPUs priced from about $38,600, why he now prefers Max-Q cards over standard workstation versions, and the heat and power tradeoffs of stacking GPUs in one machine.
📺 Source: sentdex · Published November 11, 2026
🏷️ Format: Benchmark Test
