Let's Run DeepSeek V3.2 - LOCAL AI "More Genius" than GPT-5 & Gemini 3

Value of the Information & Strength of the Argument

The video offers significant hands-on value through real-world demonstrations of running a large AI model locally. The host’s comparisons between quantizations and his explanation of token-level influences provide actionable knowledge. However, the argumentation is based on a limited set of ad-hoc tests (e.g., a few logic puzzles), which may not generalize. The host does not provide systematic benchmarks, so the claims about ‘genius’ performance are supported only by anecdotal evidence. Nevertheless, the transparent discussion of tool-calling behavior and token sampling techniques gives useful practical guidance.

Scientific Rigor, Source Quality, Title Accuracy

The host uses official model files from Hugging Face and a third-party inference app (Inferencer), which adds credibility. He does not cite academic sources or systematic evaluations. The title overstates the comparison with GPT-5 and Gemini 3, as those models are not directly tested in the video. The content is more about the local running experience than a definitive performance comparison. Despite this, the video provides clear documentation of the process and openly discusses limitations, such as the language bias issue and the speed variations.

The title claims superiority over GPT-5 and Gemini 3, but the video only tests a few reasoning puzzles and does not directly compare with those models. However, it does demonstrate the model's capabilities, so it is moderately aligned.

7/10

The video provides hands-on demonstrations and practical insights, but the testing methodology is anecdotal without rigorous benchmarks. The host is a developer, not a researcher, and relies on personal observations.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction to DeepSeek V3.2 and overview of the test plan.
• 01:30Launching DeepSeek on Mac Studio with Q5 quantization, discussing token speeds.
• 02:30Surgeon puzzle test: comparing Q5, Q6, and Speciale reasoning outcomes.
• 05:00Setting up distributed compute for Q6 quant across MacBook Pro and Mac Studio.
• 07:30Trolley problem test and reasoning comparison between versions.
• 10:00Demonstration of tool calling with webpage content retrieval.
• 12:30Testing Speciale with tool calls and a cryptographic puzzle.
• 15:00Conclusion and final thoughts on the model's capabilities.

• Inferencer App — The host uses Inferencer v1.7.3 to run the models locally.
• DeepSeek-V3.2-MLX-5.5bit (Hugging Face) — Model file for the Q5 quantization used on Mac Studio.
• DeepSeek-V3.2-Speciale-MLX-5.5bit (Hugging Face) — Model file for the Speciale deep-thinking variant.
• Kimi K2 Thinking — Companion video comparing another AI model.
• OpenAI GPT-OSS Review — Companion video reviewing OpenAI's open-source model.
• Qwen 3.1 Review — Companion video reviewing Qwen 3.1.
• Mac Studio Review — Companion video reviewing the Mac Studio hardware used in this video.