I Ran DeepSeek V4.1 at MAX on a Local AI Cluster (It’s Terrifying)

Item: I Ran DeepSeek V4.1 at MAX on a Local AI Cluster (It’s Terrifying)
Rating: 4
Author: PostSingularity Institute

by xCreate · View the channel on PSI

Applied Sciences & Engineering AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence

🎙 xCreate 👥 27K 📅 October 5, 2026 ⏱ 23 min 👁 12K 📄 expert opinion 🧭 2026-10-15

Available in: English (current) Français

Keywords

DeepSeek V4.1local AI clusterthinking tokenscode generationLLM quantization

Summary

In this video, the creator tests DeepSeek V4.1 Flash Edition locally at maximum thinking levels. He uses various quantization versions and runs the model on a cluster of two Macs (M3 Ultra and M4 Max). He presents 12 different tests, including simulating a nuclear explosion, generating a photorealistic human face, recreating games like Age of Empires II, GTA V, DOOM, Call of Duty, Street Fighter, Tomb Raider, Assassin’s Creed II, and Gears of War, as well as building a Blender-like 3D editor and an iMovie-like video editor. The creator analyzes the impact of thinking tokens (1k, 10k, 25k, unlimited) on output quality, noting that more tokens do not always mean better results. He also discusses runtime errors, loop detection, and the performance of a clustered setup. The video concludes that DeepSeek V4.1 is impressive but requires careful prompt engineering and may produce inconsistent results.

144 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides valuable insights into running large language models locally, especially the effect of thinking tokens and quantization. The arguments are supported by real demonstrations, though they are anecdotal. The creator compares different settings and identifies potential issues, contributing practical knowledge for enthusiasts and professionals.

Scientific Rigor, Source Quality, Title Accuracy

The creator relies on his own tests and does not cite independent academic sources. He provides links to his own quantization files, which are not peer-reviewed. The title uses hyperbole but is generally accurate. The video is more of a hands-on showcase than a rigorous scientific study.

108 words

Title / Content Match

The title accurately reflects the content: the creator runs DeepSeek V4.1 at maximum thinking on a local AI cluster, and the impressive results justify the 'terrifying' hyperbole.

Quality & Reliability

5/10

The video provides a hands-on demonstration of DeepSeek V4.1 with various scenarios, but the evaluation is subjective and lacks rigorous statistical analysis. Claims about model capabilities are based on personal tests, and the presence of affiliate links may introduce bias.

Chapters

Cited Sources

DeepSeek-V4.1-MLX-Q4i — Quantization file used for testing various thinking token levels.
DeepSeek-V4.1-MLX-Q9 — Higher fidelity quantization used for the clustered setup.
Inferencer — Tool used for running the inference.
GLM 5.3 Flash — Companion video on another model.
Qwen 4 Exp — Companion video on another model.
Kimi K3 — Companion video on another model.
GLM 5.3 — Companion video on another model.

External References

vtudio.com

Contribution & Novelties

This video contributes a comprehensive practical evaluation of DeepSeek V4.1, highlighting real-world performance on local hardware. It demonstrates scalable inference via clustering and the importance of thinking tokens. Compared to typical benchmark reports, this provides direct visual outputs. For further reading:

Large language model — essential background for the model tested.
Deep learning — fundamental technology behind the model.
Quantization (signal processing) — conceptual basis for the Q4i/Q9 quantizations.
Ollama — popular tool for local LLM inference, relevant to the setup shown.

82 words

Radar Profile

The radar profile shows moderate scores across all dimensions, with a slight emphasis on information quantity. The low reliability score reflects the subjective nature of the tests, while technical level is moderate, suitable for a general tech audience.

Reliability 5/10