I Ran GLM 5.3 at MAX Locally for 200+ Hours... Forget GPT-6 & Fable 🤯

I Ran GLM 5.3 at MAX Locally for 200+ Hours... Forget GPT-6 & Fable 🤯

by xCreate · View the channel on PSI

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 xCreate 👥 26K 📅 September 9, 2026 ⏱ 50 min 👁 4K 📄 original study 🧭 2026-09-09
Available in: English (current) Français

Keywords

GLM 5.3Local AIThinking modesToken efficiency3D generation

Summary

This video reviews GLM 5.3, an open-weight language model, running locally on a Mac Studio. The creator introduces a ‘MinMax thinking mode’ designed to achieve maximum intelligence with minimal token consumption, addressing the model’s tendency to over-think and generate excessive reasoning tokens. Through a series of complex prompts, including 3D solar systems, nuclear bomb simulations, a controllable spacecraft, and a procedural human face, the creator compares Low, High, Max, and MinMax thinking modes. Token counts and output quality are analyzed, showing that MinMax often delivers results comparable to Max at a fraction of the tokens (e.g., 12,000 vs. 74,000 for a space scene). The video also explores reasoning leakage, threshold-based thinking limits, and quantization artifacts. While MinMax does not always outperform Low thinking, it generally provides better quality and can be customized. The creator concludes that local GLM 5.3 is usable, especially with MinMax, but acknowledges bugs and errors, likely due to quantization or long-context limits.

156 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video offers valuable empirical data on token usage and output quality across different thinking modes, which is useful for practitioners running large models locally. The argumentation is convincing: the creator systematically compares scenarios, provides exact token counts, and includes side-by-side demonstrations. The MinMax concept is well-illustrated with concrete examples, and the discussion of reasoning leakage is insightful. However, the evaluation is subjective, relying on visual inspection rather than quantitative metrics. The creator also notes that some differences are marginal, and results vary by task. The argument that MinMax saves tokens without sacrificing much quality is supported by several examples, but the lack of statistical rigor and controlled conditions weakens the overall strength of the claims.

Scientific Rigor, Source Quality, Title Accuracy

The video is based on original testing rather than citing external sources, which is appropriate for a hands-on review. The description includes links to the model (HuggingFace), the inference app (Inferencer), and companion videos, but no scientific references. The methodology is not strictly controlled—only one run per mode is shown, and quantisation may influence results. The title is catchy and partially sensational, but it accurately reflects the extensive testing undertaken. The content aligns with the title: the creator did run GLM 5.3 at max for many hours and explored thinking modes. However, the title’s reference to ‘Forget GPT-6 & Fable’ is not substantiated in the video, making it somewhat misleading.

241 words

Title / Content Match

The title is engaging and accurately reflects the video's focus on exhaustive local testing of GLM 5.3, though the mention of GPT-6 and Fable is sensational.

Quality & Reliability

7/10

Hands-on empirical testing with detailed token counts and qualitative comparisons, but lacks rigorous methodology and may be affected by quantization artifacts and uncontrolled variables.

Chapters

Cited Sources

External References

Contribution & Novelties

The video introduces and demonstrates a ‘MinMax thinking mode’ for local LLMs, which aims to reduce token usage while preserving high intelligence. This is a practical contribution for users of large open-weight models on consumer hardware, as it addresses the common problem of excessive reasoning overhead. The creator also explores threshold-based thinking limits and discusses reasoning leakage, offering strategies to control model behavior. The empirical comparisons across multiple complex prompts provide valuable insights into token-quality trade-offs.

Pour aller plus loin :

  • GLM-5.3-MLX-Q4i on Hugging Face — The exact model quantisation used, relevant for reproduction.
  • Reasoning Leakage — Concept discussed in the video; no reliable URL available, but related to controlling chain-of-thought generation in LLMs.
  • Quantization (MLX) — The technique used to reduce model size; see Apple MLX documentation (no direct URL provided).

132 words

Radar Profile

The profile shows strong performance in information quantity and technical depth, with slightly lower scores in reliability and information quality. This suggests the video is rich in hands-on data but lacks formal scientific rigor, making it a useful practical guide rather than a peer-reviewed analysis.

Reliability 6/10