What This Is
A Reddit developer linked four NVIDIA DGX Spark desktop AI workstations into two paired clusters and benchmarked four models — DeepSeek V4 Flash, Qwen3.8 Flash Next, Qwen3.8-27B, and Qwen3.6-35B-A3B — zeroing in on how "long context" (up to 900,000 tokens) and "varying thinking depths" affect outcomes.
The standout finding in his published data: switching AI into "deep thinking" mode (thinking level set to high or xhigh) actually pushed scores below low mode while tripling the output token count. The most extreme case: DeepSeek's high-thinking mode took 2.5x longer than low mode, yet scores fell from 20/24 to 16/24.
Industry View
This contradicts the dominant AI marketing narrative. Over the past year, mainstream vendors have positioned "thinking" and "deep reasoning" as premium features — OpenAI's o-series, Anthropic's extended thinking, and various domestic "deep thinking" toggles are all priced several times higher than standard mode, on the premise that "thinking longer yields better answers."
The tester's verdict: on his task set (12 code-generation + 12 code-completion problems), deep thinking is "actively counterproductive." He recommends low or off for daily work and reserving higher settings for genuinely complex reasoning.
We should flag the counterpoints too: this is a single test series (24 problems), not an official LiveBench leaderboard result — the sample is small; tasks skew toward code completion, so the conclusion may not generalize to math competitions, academic writing, or other harder domains. We doubt model vendors will accept this finding. Treating it as "one user, one scenario" is the prudent read.
Impact on Regular People
For enterprise IT: if your company is buying AI inference APIs, the "deep thinking on or off" toggle is a direct cost variable. Leaving it on by default is like paying 3x the token price — whether that money is well spent is a business-side judgment call.
For individual professionals: in daily AI assistant use, try flipping off the "deep thinking" button and compare speed against answer quality. "Looking more professional with it enabled" isn't necessarily true.
For the consumer market: desktop workstations like the DGX Spark can run 900,000-token contexts, which tells us the hardware bar for running large models locally keeps dropping — costs for individuals and small companies building their own AI toolchains will continue to fall.