QWEN 3.6 35B sur 4GB VRAM ??! : La meilleure configuration

Value of the Information & Strength of the Argument

The video provides concrete, actionable advice for running a large model on constrained hardware. The author clearly explains the technical rationale for each parameter (e.g., why Q4 is better than Q3, why UBatch should be 512, why n-gram mode helps). He demonstrates the setup with real measurements and shows how to tweak settings to balance performance. The argumentation is empirical, based on his testing, and he acknowledges the limitations of a 4GB GPU, such as slower pre-fill. However, some comparisons (e.g., with Gemma 4) are made without supporting data, and the claims about the model’s quality rely on his own observations rather than external benchmarks.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is moderate: the content is a practical tutorial rather than a formal study. The author shares his own results and settings but does not reference external sources or benchmarks to substantiate claims about model performance or superiority. The only linked source is his own ‘Jean’ tool, which is not a research reference. The title accurately describes the content, and the video stays on topic. The lack of citations reduces the overall rigor, but the methodology (test, adjust, measure) is sound.

The title accurately reflects the content: it promises showing how to run a 35B model on 4GB VRAM and the video delivers exactly that, with configuration steps and performance tests.

8/10

The tutorial is based on hands-on testing with a real GTX 1650 4GB GPU, showing measured tokens per second. The author explains the reasoning behind each setting (quantization, context, KV cache, etc.) and addresses limitations. Claims about Qwen 3.6 being better than Gemma for agents are presented as the author's own assessment without external benchmarks, which slightly lowers the score.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:30Introduction: running Qwen 3.6 35B on a GTX 1650 with only 4GB VRAM.
• 02:00Explanation of the A3B architecture and why it fits in low VRAM; choice of Q4_K_M quantization.
• 05:00Detailed configuration: context 100k, KV cache Q4, threads, UBatch 512, ngl 999, disabling reasoning.
• 07:00Advanced flags: n-cpu-moe=39, flash attention, mlock, n-gram mode, and session handling.
• 08:00First benchmark: generation speed ~23-24 tokens per second, pre-fill ~55 tokens per second.
• 09:30Tuning for better pre-fill: UBatch 1024, n-cpu-moe=40, results at ~83 tokens per second.

• Jean - graphical interface for llama.cpp — The author's own tool, used in the video to display and modify model configurations, which he plans to release open-source.