DeepSeek V4 : le modele qui humilie les IA américaines

Item: DeepSeek V4 : le modele qui humilie les IA américaines
Rating: 4
Author: PostSingularity Institute

by Eliott Meunier

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence

🎙 Eliott Meunier 👥 51K 📅 June 6, 2026 ⏱ 35 min 👁 5K 📄 expert opinion 🧭 2026-08-27

Available in: English (current) Français

Keywords

DeepSeek V4RoPEMLACSAHCAMixture of ExpertsKV cacheopen sourceefficiencyHuawei GPUs

Summary

The video presents DeepSeek V4, a new open-source LLM with 1.6 trillion parameters, trained on Huawei GPUs by a team 40-100 times smaller than OpenAI’s. The author highlights its impressive performance on benchmarks, rivaling models like Opus and GPT-5, while being significantly cheaper (up to 100x less). He then explains the key technical innovations that enable this efficiency: RoPE (rotary position embedding) for better positional encoding, MLA (Multi-head Latent Attention) which compresses the KV cache by 96%, CSA (Cross-layer Sparse Attention) with a Lightning Indexer that selects only relevant context blocks, HCA (Hierarchical Compressive Attention) that compresses tokens into blocks for a global view, and Mixture of Experts with 256 experts, activating only 9 per token. The author also demonstrates the model’s capabilities with a practical example of creating a website, comparing V4 Flash, V4 Pro, and Opus 4.7. He emphasizes the cost-effectiveness and open-source nature of the model, suggesting it is a major step for the AI community.

159 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides a high-value, in-depth explanation of DeepSeek V4’s architecture, going beyond surface-level benchmarks to explain the underlying innovations. The author’s argumentation is logical and well-structured, systematically comparing each innovation to the original 2017 Transformer. He uses clear analogies (e.g., dictionary definitions, rotating vectors) to make complex concepts accessible. The practical demonstration of creating a website adds concrete value, showing real-world performance and cost differences. However, the argumentation is largely based on the author’s interpretation of the paper, and he does not provide direct citations or external verification, which slightly weakens the overall argument.

Scientific Rigor, Source Quality, Title Accuracy

The video demonstrates a good understanding of the technical subject, but the rigor is limited by the lack of direct references to the DeepSeek paper or other primary sources. The author mentions reading the paper but does not cite specific sections or provide links. The description includes a link to OpenRouter (a platform to use the model) and a previous video on LLMs, but these are not scientific sources. The title is somewhat clickbait (‘humiliates American AIs’), but the content is substantive and matches the title’s promise of technical analysis. The author’s personal expertise and the structured presentation contribute to a generally reliable, though not fully verifiable, analysis.

218 words

Title / Content Match

The title is somewhat sensationalist ('humiliates American AIs'), but the content does focus on DeepSeek V4's technical innovations and performance, making it largely appropriate.

Quality & Reliability

7/10

The video provides a detailed technical breakdown of DeepSeek V4's architecture, based on the author's reading of the paper. The explanations are clear and structured, but the analysis is subjective and lacks direct citations to the paper or external sources. The author's expertise is evident, but the lack of verifiable references and the promotional tone for his own content reduce the overall reliability.

Chapters

Cited Sources

OpenRouter — Mentioned as a platform to use DeepSeek models with zero data retention option.
How LLMs work (video by Eliott Meunier) — Referenced as a prerequisite for understanding the basics of LLMs.

Concurring Sources

DeepSeek-V3 Technical Report — The technical report for DeepSeek V3, which shares many architectural innovations (MLA, MoE) with V4, providing a basis for the claims made in the video.

Dissenting Sources

No direct discordant sources found — The video's claims are not directly contradicted by any known sources, but the lack of citations makes it difficult to verify all details.

Contribution & Novelties

The video offers a clear, accessible explanation of DeepSeek V4’s novel architectural innovations, particularly the combination of MLA, CSA, and HCA for efficient attention, and the use of a Lightning Indexer for sparse attention. It highlights the significance of these innovations in enabling a 1.6 trillion parameter model to run efficiently on less powerful hardware. The author’s practical demonstration of cost and performance adds original value beyond just theoretical discussion.

Pour aller plus loin :

Rotary Position Embedding (RoPE) — The original paper introducing RoPE, which is a key innovation discussed in the video.
Multi-head Latent Attention (MLA) — The paper describing MLA, the attention mechanism that compresses the KV cache, as used in DeepSeek V2 and later.
Mixture of Experts (MoE) — The foundational paper on MoE, which is a core component of DeepSeek V4’s architecture.

137 words

Radar Profile

The radar profile shows high scores in quantity of information, quality of information, and technical level, indicating a content-rich and technically detailed video. The lower score in global reliability reflects the lack of direct citations and the subjective nature of the analysis.

Reliability 6/10