I Ran DeepSeek V4.1, Qwen & GLM on an Old M1 Laptop... It Worked! 🤯

I Ran DeepSeek V4.1, Qwen & GLM on an Old M1 Laptop... It Worked! 🤯

Title history
  • I Ran DeepSeek V4.1, Qwen & GLM on an Old M1 Laptop... It Worked! 🤯 — observed on 20261029-0033 UTC (YouTube)

by xCreate · View the channel on PSI

Applied Sciences & Engineering AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 xCreate 👥 27K 📅 October 28, 2026 ⏱ 15 min 👁 237 📄 tutorial 🧭 2026-10-29
Available in: English (current) Français

Keywords

deepseek v4.1qwen 3.8 flashglm 5.3 flashssd streaminglocal ai

Summary

The video, hosted by xCreate, demonstrates running large open-weight AI models—DeepSeek V4.1 Flash, Qwen 3.8 Flash Next, and GLM 5.3 Flash—on a 5-year-old M1 Max MacBook Pro with 64GB RAM, an M5 Max with 128GB, and a Mac Studio with 512GB. The creator uses sophisticated streaming techniques (expert streaming, engram streaming, NTP) to fit models exceeding 500GB into limited memory. Initial tests on the M1 Max achieve 1.6 tokens per second for DeepSeek, rising to 5-8 tokens per second on the M5 Max. On the Mac Studio, the standard (non-streaming) version reaches 15-16 tokens per second, while the streaming edition tops out at 11 tokens per second but uses far less RAM. Qwen 3.8 Flash Next with 9-bit quantization hits ~30 tokens per second (MTP disabled) on the M5 Max, doubling to ~50 with MTP enabled. GLM 5.3 Flash runs at a mere 2 tokens per second on the 128GB system. The video concludes that streaming techniques enable frontier-level intelligence on older, memory-limited hardware, though native loading remains faster when RAM permits. The creator also hints at future tests with Kimi K3.1 and plans to investigate prefetching for further speed gains.

191 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video offers hands-on, quantitative evidence of running massive AI models on consumer-grade hardware, which is valuable for hobbyists and researchers exploring local inference. The argumentation is based on direct observation—token speeds, memory usage, and side-by-side comparisons across three systems. However, the analysis lacks scientific rigor: there is no controlled benchmark methodology, no repetition for statistical significance, and the effects of different quantization levels and streaming parameters are reported anecdotally. The creator acknowledges some limitations (e.g., thrashing effects, SSD wear) but does not systematically isolate variables. Overall, the information is useful as a practical demonstration, but conclusions about optimal techniques should be taken with caution.

Scientific Rigor, Source Quality, Title Accuracy

Scientific rigor is moderate: the creator references specific tools (Inferencer) and companion videos but provides no academic sources or formal documentation for the streaming techniques (engram streaming, expert streaming). The measurements appear honest (e.g., noting when speeds drop due to expert eviction), yet the lack of controlled conditions limits reproducibility. The title is accurate and not exaggerated—it matches the content’s thrust. The video’s strength lies in its practical, real-world testing rather than theoretical depth. Adequation between title and content is strong, with no clickbait beyond the enthusiastic exclamation.

208 words

Title / Content Match

The title accurately reflects the content: the creator successfully runs large open-weight models on a 5-year-old M1 laptop using streaming techniques. The exclamation adds subjective excitement but does not mislead.

Quality & Reliability

6/10

The video provides concrete performance benchmarks on real hardware, but the methodology is informal (single tester, no controlled conditions, no external validation). Claims about streaming techniques are plausible but not backed by peer-reviewed sources. Reliability is moderate.

Chapters

Cited Sources

External References

Contribution & Novelties

The original contribution is demonstrating that cutting-edge open-weight models (DeepSeek V4.1, Qwen 3.8, GLM 5.3) can run on aging hardware via intelligent streaming of experts and n-grams, achieving usable token rates. The video provides concrete performance numbers across three Apple Silicon configurations, highlighting the trade-offs between RAM capacity and quantisation levels. It also introduces the concept of ‘streaming editions’ (suffixed ‘se’) and the importance of NTP for speed. The results challenge the assumption that large models require high-end GPUs or cloud infrastructure.

Pour aller plus loin :

  • Mixture of experts — Core technique enabling selective activation of model weights, relevant to expert streaming.
  • Quantization (signal processing) — Explains how reducing bit-width (e.g., 9-bit vs 6-bit) impacts model performance and memory footprint.
  • Solid-state drive — The storage medium used for offloading model weights; crucial for understanding I/O bottlenecks.

137 words

Radar Profile

The radar profile shows strong scores in quantity of information and technical level, reflecting the detailed practical measurements. However, quality of information and global reliability are lower, indicating that the content is informative but lacks methodological rigor and external validation. The overall impression is that of an enthusiastic, hands-on exploration rather than a rigorous scientific study.

Reliability 5/10