I Ran DeepSeek V4.1 Locally for 24 Hours... Local AGI is HERE!? 🤯

Item: I Ran DeepSeek V4.1 Locally for 24 Hours... Local AGI is HERE!? 🤯
Rating: 3
Author: PostSingularity Institute

by xCreate · View the channel on PSI

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQDArtificial general intelligence (AGI)

🎙 xCreate 👥 26K 📅 September 11, 2026 ⏱ 28 min 👁 2K 📄 expert opinion 🧭 2026-09-11

Available in: English (current) Français

Keywords

DeepSeeklocal LLMbenchmarkquantizationAI testing

Summary

The creator runs DeepSeek V4.1 locally on a Mac Studio M3 Ultra, using a quantized MLX build. They begin by discussing the model architecture, noting the retirement of V4 Pro and the addition of engrams, simplified sparse attention, and about 16B active parameters during decode. They then run a series of tests covering math, coding, logic, office applications (Photoshop, PowerPoint, Word), canvas animations, 3D city simulations, racing games, flight simulators, and more. They highlight the model’s strong performance on DeepSWE benchmarks, claiming it beats Opus and GLM 5.3 in coding tasks. They also discuss temperature mechanics, showing that with temperature zero the model gives correct answers on math problems that previously failed. They implement and explain ’thinking limits’ — a custom feature that caps the number of reasoning tokens — which trades off some quality for speed. The video shows many generations, some impressive (like a go-kart game, FPS zombie shooter, and a photo-realistic face) and some failures (like GTA and a 3D city). The creator notes that for complex prompts, thinking is required, but thinking modes consume huge token budgets. They conclude that the model is a frontier open-weight model with an MIT license, but note that it sometimes lacks personality and fails on certain logic questions. Overall, they recommend it for local use with thinking limits, especially for coding and general productivity tasks.

225 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides valuable hands-on experience of running a large language model locally, offering concrete insights into quantization, temperature settings, and token budget management. However, the testing is unstructured and subjective, with no formal methodology or statistical analysis. The creator often makes speculative claims (e.g., about architecture improvements) without external verification. The argumentation is primarily anecdotal, based on individual prompts rather than systematic evaluation. Despite this, the demonstration of practical usage scenarios and the introduction of ’thinking limits’ as a workaround for resource constraints is a useful contribution for other local AI enthusiasts.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is limited; no academic sources are cited, and the creator relies on the model’s own benchmarks and personal tests. The sources used are the HuggingFace page for the quantized model and the Inferencer test system, which are directly relevant. The title accurately reflects the content, though it uses sensational language (‘Local AGI is HERE!?’) that overstates the model’s capabilities. The video includes several affiliate links, but they are clearly marked and do not influence the technical content. No comments were provided, so public reception cannot be assessed.

198 words

Title / Content Match

The title accurately describes the content: the creator runs DeepSeek V4.1 locally and discusses its capabilities and potential.

Quality & Reliability

5/10

Hands-on testing with personal observations; results are subjective and based on a quantized build, with benchmarks partly reliant on the model's own reporting.

Chapters

Cited Sources

DeepSeek-V4.1-MLX-Q4i on HuggingFace — Quantized model used for local testing
Inferencer — Test system used for running the model
GLM 5.3 Flash Video — Companion video comparing with another model
Qwen 4 Exp Video — Companion video comparing with another model
Kimi K3 Video — Companion video comparing with another model
GLM 5.3 Video — Companion video comparing with another model

External References

vtudio.com

Contribution & Novelties

The video offers a practical, in-the-wild evaluation of DeepSeek V4.1 on a high-end Mac, highlighting the feasibility of running a 552B-parameter model locally with quantization. The introduction of ’thinking limits’ as a custom token budget control is a novel workaround that balances quality and speed for local inference. The creator also demonstrates the model’s strengths in coding and office application generation, while noting weaknesses in complex interactive simulations.

Pour aller plus loin :

Mixture of experts — Explains the architecture used in many large language models, including DeepSeek.
Quantization (signal processing) — Covers the fundamentals of quantization, essential for reducing model size for local deployment.
Large language model — Provides background on LLMs and their capabilities.

116 words

Radar Profile

The radar shows moderate scores across most dimensions, with relatively higher technical depth and information quantity, but lower reliability and quality, reflecting the anecdotal and subjective nature of the testing.

Reliability 5/10