Qwen 3.8 - Local AI xHigh Thinking is SCARY 🤯

Qwen 3.8 - Local AI xHigh Thinking is SCARY 🤯

by xCreate · View the channel on PSI

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 xCreate 👥 26K 📅 August 17, 2026 ⏱ 25 min 👁 14K 📄 science communication 🧭 2026-09-10
Available in: English (current) Français

Keywords

Qwen 3.8High ThinkingLoop DetectionQuantizationToken Inspector

Summary

The video reviews Qwen 3.8, a 27-billion-parameter open-source language model, running locally on a Mac Studio with the Inferencer app. The creator, xCreate, thoroughly tests the ‘Extra High Thinking’ mode across various prompts, comparing it to lower thinking settings. He demonstrates procedural generation (faces, 3D anatomy) and quick game clones (GTA, Red Dead, Final Fantasy, Mario, Outrun), highlighting both impressive results and frequent thinking loops. The video explains technical aspects like token entropy, temperature, quantization (Q4, Q9, Q10), and introduces a ’thinking detector’ to break loops. Despite errors and token inefficiency, the model is praised for its capabilities at 27B parameters, even beating some cloud models. The presenter shares tips for managing loops and improving output, concluding that Qwen 3.8 is a strong start for local AI projects, though it may require iterative prompting and error fixing.

137 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides high value through extensive hands-on testing and clear explanations of AI inference mechanics. The creator demonstrates a systematic approach—testing different quantizations, temperatures, and inference engines—to isolate issues like thinking loops. The argumentation is strong when discussing technical causes (e.g., quantization errors, entropy randomness) and practical solutions (loop detection, low temperature). However, some claims are anecdotal and lack statistical backing, and the generalization from a few game demos to overall model quality is optimistic. The comparison with Google Gemini Pro is illustrative but brief. Overall, the evidence is compelling for the model’s creative potential, though the reliability of long outputs is questioned.

Scientific Rigor, Source Quality, Title Accuracy

The video exhibits moderate scientific rigor: it is transparent about failures, shows raw outputs, and explains tools used, but lacks controlled benchmarking and external references. Sources are primarily the model’s Hugging Face page and the Inferencer app, with no scholarly citations. The title is somewhat clickbaity (‘SCARY’) but aligns with the impressive yet unsettling capabilities demonstrated. Public comments (29 analysed) are predominantly positive, with many users sharing their own loop experiences and suggesting causes (MLX vs. llama.cpp, chat templates). There is no evidence of malicious criticism; the debate is technical and constructive.

211 words

Title / Content Match

Title is slightly sensational but accurately reflects the focus on High Thinking mode and its impressive yet glitchy outputs.

Quality & Reliability

7/10

The video presents original empirical tests of a local LLM, with detailed explanations of technical concepts. However, it is a single-source anecdotal review without peer review or external validation, and the creator acknowledges issues (looping, errors) that may be environment-specific.

Chapters

Cited Sources

External References

Contribution & Novelties

The video offers a unique hands-on evaluation of Qwen 3.8’s ‘High Thinking’ mode, particularly highlighting the phenomenon of thinking loops and a practical detection tool. It contributes to the local AI community by demonstrating the trade-offs between quantization levels and reasoning quality, and by explaining advanced concepts like token entropy and multi-token prediction. The creator’s ’thinking loop detector’ is an original contribution that helps mitigate a common issue.

Pour aller plus loin :

  • Quantization for LLMs — Explanation of how quantization reduces model size and affects output quality.
  • MLX (Apple) — Framework for efficient machine learning on Apple silicon, used for the quantized models.
  • Speculative decoding — Technique mentioned in the video for accelerating inference, relevant to the MTP discussion.

120 words

Radar Profile

The radar profile shows high scores in quantity and technical depth, but lower reliability. This indicates a detailed, informative video with solid explanations, yet the lack of external validation and reliance on anecdotal evidence lowers its trustworthiness for rigorous scientific use.

Reliability 7/10

💬 Globalement positif, avec de nombreuses discussions techniques sur les causes des boucles (MLX vs. llama.cpp, templates de chat). Sur les 29 commentaires analysés, la majorité félicite la démonstration et les explications, certains partagent leurs propres expériences de performance, et un débat constructif s'engage sur la reproductibilité sans hostilité.