Qwen 3.8 Flash via SSD: Max Quants on Low RAM (vs Kimi 3.1 & Qwen 4) 🤯

Qwen 3.8 Flash via SSD: Max Quants on Low RAM (vs Kimi 3.1 & Qwen 4) 🤯

by xCreate · View the channel on PSI

Applied Sciences & Engineering AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 xCreate 👥 27K 📅 October 23, 2026 ⏱ 29 min 👁 207 📄 tutorial 🧭 2026-10-23
Available in: English (current) Français

Keywords

quantizationstreamingmemory optimizationMTPcontext attention

Summary

This video by xCreate explores running Qwen 3.8 Flash (based on Qwen 4 EXP architecture) on Mac computers with limited RAM by streaming N-Gram weights from SSD storage. The creator compares Q4, Q6, and Q9 quantizations, demonstrating that higher quant levels (Q6, Q9) can run with lower memory usage than Q4 when using streaming, while delivering better quality generations. Tests include 3D scenes like volcano explosions, motorbike racing, and helicopter assembly, all generated locally with variable token counts. The video also covers processing intensity adjustments, thermal throttling on MacBook Pro, and the impact of context attention on quality and token counts. Key findings: Q6 streaming uses less memory than Q4 and produces better results; Q9 provides highest quality but requires more VRAM; enabling MTP improves speed. The presenter also discusses token probability and quantization errors, and shows how disabling sparse attention (context attention) reduces hallucinations but increases memory and computation. Overall, the video demonstrates that with streaming, high-quality LLM inference is feasible on moderate hardware, and compares the results with upcoming Kimi K3.1 and Qwen 4.

176 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides concrete performance metrics (tokens/sec, memory usage) and visual comparisons across quantizations, which is valuable for anyone attempting to run large local LLMs. The argumentation is based on direct observation and hands-on trials, not just speculation. The creator explains the trade-offs between quantization levels, MTP, and context attention in a practical way. However, the lack of a controlled experimental design (varying prompts, multiple runs, statistical analysis) weakens the strength of the conclusions. The demonstrations are convincing but not systematically reproducible.

Scientific Rigor, Source Quality, Title Accuracy

The video references specific Hugging Face model repositories for the streaming quantizations (Q6s, Q9s) and the Inferencer tool, which are legitimate sources. The companion videos cited are from the same channel and not external scientific references. The title accurately describes the main theme and does not overpromise. However, the video does not cite any academic papers or official documentation, relying solely on the creator’s empirical results. The methodology is transparent enough for enthusiasts, but the absence of a rigorous benchmark protocol reduces scientific credibility. No external sources are used to cross-validate the findings.

190 words

Title / Content Match

The title accurately reflects the content: the video tests Qwen 3.8 Flash with maximum quantization levels running on a limited-RAM Mac using SSD streaming, and compares performance with Kimi 3.1 and Qwen 4 benchmarks.

Quality & Reliability

6/10

The video presents real hands-on testing with concrete memory and speed figures, but lacks rigorous methodology, controlled conditions, and external validation. The author demonstrates expertise but the results are anecdotal and hardware-specific.

Chapters

Cited Sources

Concurring Sources

  • DeepSeek 4.1 Streaming — Same channel's prior video demonstrating streaming engrams for a larger model.
  • Qwen 3.8 Flash — Companion video testing the same model without streaming.

External References

Contribution & Novelties

The video introduces a novel technique for running high-quantization LLMs on low-RAM systems by streaming N-Gram tables from storage, claiming no performance loss. It provides practical benchmarks and visual quality comparisons across Q4, Q6, and Q9, showing that Q6 streaming outperforms Q4 in both speed and quality while using less memory. The creator also demonstrates the trade-offs of disabling sparse attention (context attention) to reduce hallucinations.

Pour aller plus loin :

  • Mixture of experts — The underlying architecture (Qwen 4 EXP) uses MoE, which is relevant to understanding the N-Gram streaming concept.
  • Large language model — Provides background on LLM architecture and inference challenges.
  • Model compression — Quantization is a form of model compression; this gives theoretical context.
  • N-Gram streaming — A technique emerging for streaming model weights; no reliable external reference found for this specific term.

137 words

Radar Profile

The profile shows a highly technical video with strong depth in technical execution (niveau_technique=9) but weaker on methodological reliability (fiabilite_globale=6). The quantity and quality of information are adequate, indicating a well-packed demonstration but not a scientifically rigorous evaluation.

Reliability 6/10