Cette IA chinoise est 17x moins chère que GPT et Claude... et elle les BAT ?

Value of the Information & Strength of the Argument

The video provides valuable information about a significant open-source AI release, detailing its architecture, benchmarks, and cost implications. The argumentation is structured and generally balanced, acknowledging both strengths and weaknesses. The creator effectively explains technical concepts like mixture-of-experts and quantization in an accessible manner. However, the analysis relies heavily on the creator’s interpretation of benchmark results and does not critically examine the methodology behind the claimed 80% superiority. The promotional segment for the creator’s training program, while clearly separated, somewhat detracts from the objective tone.

Scientific Rigor, Source Quality, Title Accuracy

The video cites specific benchmarks (OmniDocBench, BrowserComp, SWE-bench) and technical details, but does not provide direct links to primary sources or official documentation. The description contains only links to the creator’s own newsletter and training program, not to Alibaba’s release notes or benchmark papers. The title is somewhat clickbait, but the content does address the claims with nuance, giving a fair assessment. The overall scientific rigor is moderate; the information appears accurate but is presented without independent verification.

The title is somewhat sensationalist ('17x cheaper', 'BAT ?') but the content does address these claims, providing a nuanced verdict that Qwen 3.5 is not universally superior but offers a strong cost-performance ratio.

6/10

The video provides a balanced overview of Qwen 3.5's capabilities and limitations, citing specific benchmarks and technical details. However, it relies heavily on the creator's interpretation and promotional claims, without independent verification or primary sources. The presentation is clear but lacks depth in critical analysis of benchmark methodology.

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

• 00:00Introduction: Qwen 3.5 release and claims of beating GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro.
• 00:45Architecture: native multimodality with 'grand fusion' training on text, images, and video.
• 02:00Technical innovation: Gated Delta Networks, linear attention, 19x faster decoding, 256k context.
• 04:00Benchmark analysis: strengths in document recognition and agentic search, weaknesses in math and coding.
• 06:00Cost-performance: 17B active parameters, 10-17x cheaper than GPT/Claude, 8x more efficient.
• 08:00Local deployment: 807GB full model, quantization to 214GB (4-bit) or 192GB (3-bit), requires high-end hardware.
• 10:00Language support: 2011 languages, 250k vocabulary, 15-40% fewer tokens for non-Latin scripts.
• 11:00Verdict: Qwen 3.5 is not the best in all areas but is the most complete open-source model, with strategic positioning in agentic AI.

• Vision IA Newsletter — Creator's newsletter for AI news summaries.
• Vision IA Training Program — Creator's paid AI training course, mentioned in the promotional segment.

• Qwen official blog — Official announcement of Qwen 3.5, likely containing benchmark details and technical specifications.

• Independent benchmark evaluations — The video claims Qwen 3.5 beats GPT-5.2 on 80% of benchmarks, but independent evaluations may show different results depending on the benchmark suite and methodology.