I Ran the FULL DeepSeek V4.1 (1017) via SSD for 24 Hrs... New #1 Local AI 🤯
Keywords
Summary
116 words
Critical Evaluation
Value of the Information & Strength of the Argument
The video provides substantial hands-on value through multiple test scenarios, directly demonstrating the model’s capabilities and limitations. The argumentation is well-structured: the creator systematically varies parameters (temperature, thinking tokens) and shows concrete outcomes, including runtime errors and fixes. He explains technical concepts like engram offloading and token sampling, supporting his claims with visual evidence. However, the argumentation relies heavily on personal experience rather than external validation, and some benchmark claims are presented without sources, somewhat undercutting the scientific rigor.
Scientific Rigor, Source Quality, Title Accuracy
Scientifically, the video lacks proper citations for external benchmarks and the engram paper, though it does reference the model repository and inference tools. The title’s ‘24 Hrs’ is not substantiated in the content, but the core premise of running the full model via SSD is upheld. No public comments were provided to analyze trends.
148 words
Title / Content Match
The title claims a 24-hour run, but the video does not specify duration; it does run the full model locally via SSD offloading, so the core claim is accurate.
Quality & Reliability
6/10
The video is a hands-on test by an experienced AI enthusiast, providing detailed benchmarks and comparisons. However, external claims (like SWE-bench scores) are not sourced, and there is an unverified mention of Anthropic's claim about DeepSeek proxying to Claude. Performance metrics depend on the creator's specific hardware setup.
Key Moments
Markers derived by PSI from the transcript: the creator did not define chapters.
- Introduction and benchmark discussion (SWE-bench 77, terminal bench, etc.)
- Volcanic explosion generation with thinking disabled and temperature 0
- Photorealistic human face generation test
- 3D human anatomy model with togglable organs
- Website generation attempts, showing runtime errors with temperature 1
- GTA-style game clone test, comparing temperature 0 vs 1
Cited Sources
- Inferencer - Inference tool — Mentioned as the inference platform used to run the model
- DeepSeek-V4.1-Flash-1017-MLXe on HuggingFace — Repository for the quantized model used in the test
External References
Contribution & Novelties
The video demonstrates that a 510GB model can be run locally with SSD offloading, achieving ~15 tokens per second on an M3 Ultra. It uniquely shows the effectiveness of engram offloading and provides practical insights into temperature settings for code generation. The comparison between local and cloud performance highlights trade-offs in speed and control.
Pour aller plus loin :
- DeepSeek — Official model page; relevant for benchmarking and technical details.
- MLX — Apple’s array framework used for efficient local inference; pertinent to the deployment approach.
- Quantization (machine learning) — Key technique for reducing model size, directly relevant to the quantized variants discussed.
102 words
Radar Profile
The profile shows high information quantity and moderate technicality, with a lower reliability score due to unsourced external claims. The balance suggests a hands-on review valuable for practical insights but requiring external verification for benchmarks.