DeepSeek V3 (DeepSeek)
The Chinese company DeepSeek releases DeepSeek V3, a large open-access model with a Mixture-of-Experts architecture (671 billion parameters, 37 billion activated per token), with performance comparable to the best proprietary models at a very low training cost.
Key points
- DeepSeek publishes DeepSeek V3 on December 26, 2024, a large openly available model.
- A Mixture-of-Experts architecture of 671 billion parameters, of which 37 billion are activated per token.
- Performance comparable to the best proprietary models for a very low training cost.
- It confirms the rise of low-cost open models as possible sources for AI assistants.
Analysis
DeepSeek V3 illustrates the emergence of open and economical models capable of rivaling the proprietary leaders. Their availability in open access makes them easy to integrate into a multitude of third-party tools, which further multiplies the engines likely to answer internet users and cite brands.
For visibility, this diffusion means AI answers can come from very varied sources and tools, less visible than the major consumer assistants. The consistency of brand information across the web remains the best lever to stay correctly represented whatever the underlying model.
What to do
- Do not limit yourself to the dominant assistants: include in your monitoring the third-party tools that rely on open models.
- Maintain consistent and up-to-date brand information across the whole web, the common source of all models.
- Prioritize presence on reliable and widely reused references, which feed both open and proprietary models.
Confirms the rise of low-cost open models as possible sources for AI assistants. Worth watching, as these models increasingly power third-party answer tools.