EXPLORINEER
en

Archive — CN — technode

DeepSeek upgrades V3 model with more parameters, open-source shift

DeepSeek announced the release of its upgraded DeepSeek-V3-0324 model with 685 billion parameters on March 24. The model's open-source license has shifted to MIT, following the path of the DeepSeek-R1 model. Known for cost-effectiveness, it surpasses competitor models while achieving a low training cost of $5.576 million.

Key points

  • Release date: March 24.
  • New version: DeepSeek-V3-0324 with 685 billion parameters; an increase from 671 billion in the previous model.
  • Open-source license changed to MIT, aligning with DeepSeek-R1.
  • The original DeepSeek-V3 outperformed models like Qwen2.5-72B and Llama-3.1-405B in tests.
  • Performance comparable to proprietary models like GPT-4o and Claude-3.5-Sonnet.
  • Total training cost for DeepSeek-V3 was $5.576 million, optimized through algorithms and hardware.

Why it matters

  • The shift to open-source licenses aligns with the growing trend towards transparency in AI development.
  • The performance of open-source models versus proprietary models highlights the increasing capabilities of accessible technology.
  • DeepSeek's cost-effective training model reflects current IT trends emphasizing budget efficiency in AI advancements.

DeepSeek's V3 upgrade marks a significant step in the evolution of cost-effective, high-performance AI models, maintaining competitiveness alongside proprietary counterparts while emphasizing an open-source approach.