I Ran Qwen 4 EXP Locally For 24 Hours... Local AI is the Frontier 🤯

I Ran Qwen 4 EXP Locally For 24 Hours... Local AI is the Frontier 🤯

by xCreate · View the channel on PSI

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 xCreate 👥 26K 📅 August 27, 2026 ⏱ 22 min 👁 36K 📄 expert opinion 🧭 2026-09-10
Available in: English (current) Français

Keywords

Qwen 4 EXPlocal LLMprocedural generationcoding testquantization

Summary

The video is a practical review of the newly released Qwen 4 EXP architecture, running under the name Qwen 3.8-Flash-Next, on high-end Mac hardware (M3 Ultra and M5 Max). The creator tests the model’s capabilities in procedural generation, coding, multimodal understanding, and logical reasoning. Despite noting several runtime errors and encoding bugs in the initial reference code, the model delivers impressive results, such as a detailed solar system simulation, a GTA-like game with free-roam controls, and a procedurally generated face with adjustable parameters. The creator also compares it to Kimi K3 and GLM 5.2, highlights the performance increase from 15 to 20 tokens per second after code optimizations, and notes that Q4 quantization runs at 43 tokens per second. The video includes safety and logic tests, showing the model correctly refuses harmful actions. The conclusion emphasizes that this 125-billion-parameter model (with only 6B active parameters) achieves near-frontier results locally, signaling rapid progress in open-source AI.

155 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video offers substantial value by directly demonstrating the model’s output in several complex generation tasks, providing concrete evidence of its capabilities. The strengths are its honest reporting of failures and bugs, along with the detailed breakdown of token counts and inference speeds. The argumentation is solid, based on empirical testing rather than speculation. However, the analysis is centered on a single hardware configuration and lacks a controlled comparison with other models beyond a few mentions. The improvement to the inference code is a personal contribution, which adds originality but also introduces potential bias. Overall, the information is engaging and informative, though not exhaustive.

Scientific Rigor, Source Quality, Title Accuracy

The video cites sources through the description, including links to HuggingFace model files and an app. These are legitimate references for the model and tooling, but the video does not cite any external papers or independent benchmarks, limiting the rigor. The title accurately matches the content, and no significant discrepancy is observed. The presentation is technically detailed, but the lack of formal documentation or peer-reviewed references lowers the scientific rigor score.

190 words

Title / Content Match

The title accurately reflects the content: the creator ran the model locally for 24 hours and showcased its frontier-like capabilities, though the 24-hour claim is plausible based on the described testing duration.

Quality & Reliability

7/10

The video provides extensive hands-on testing of the Qwen 4 EXP model with detailed demonstrations, but the analysis is based on a single creator's experience with potential biases. The model is new and the code is still being debugged, which limits reliability.

Chapters

Cited Sources

External References

Contribution & Novelties

The video’s original contribution lies in its hands-on evaluation of a cutting-edge open-source LLM, highlighting its realistic potential for local deployment. It uncovers specific bugs in the reference code and proposes fixes, adding practical value to the community. The demonstrations of high-fidelity procedural generation (solar system, GTA-like, face) are impressive and serve as a benchmark for other models. For further exploration, several key concepts are relevant.

Pour aller plus loin :

  • Mixture of experts — Explains the underlying architecture that allows sparse activation, relevant to the 6B active parameters.
  • DeepSeek — The video explicitly mentions following DeepSeek’s architecture pattern, providing background on the influence.
  • Quantization (signal processing) — Essential for understanding how Q4/Q9 quantizations trade off performance vs. accuracy, as shown in the video.

124 words

Radar Profile

The radar profile shows high scores in quantity and technical level, reflecting the extensive testing and deep detail, but moderate on quality and reliability due to the subjective nature of the evaluation and lack of external validation.

Reliability 6/10

💬 Sur les 30 commentaires analysés, le climat est très positif : les spectateurs expriment leur enthousiasme pour les performances du modèle et les possibilités de l'IA locale. Plusieurs posent des questions techniques sur le déploiement, tandis que quelques-uns font l'éloge de la démonstration et des résultats. Aucune critique négative significative n'est relevée.