DeepSeek Unveils Enhanced V3 AI Model with MIT License, Boosting Accessibility and Performance

DeepSeek Releases Improved V3 Model

DeepSeek, a Chinese artificial intelligence lab, has quietly rolled out an updated version of its DeepSeek-V3 large language model (LLM) with significant improvements and a new open-source license. The release, first reported by software developer Simon Willison, marks a notable advancement in the accessibility and capabilities of open-source AI models 1

1

Key Enhancements and Licensing

The latest iteration of DeepSeek-V3, dubbed V3-0324, introduces several notable improvements:

1. MIT License Adoption: The model has transitioned from a custom open-source license to the widely-used MIT License, allowing developers to use and modify the model in commercial projects with minimal restrictions 1

   1

   .

2. Improved Performance: Early benchmarks suggest that the new version outperforms its predecessor in programming tasks. A reported benchmark test showed the model achieving a score of about 60% in generating Python and Bash code, several percentage points higher than the original DeepSeek-V3 1

   1

   .

3. Hardware Efficiency: Despite its 671 billion parameters, DeepSeek-V3 only activates about 37 billion when responding to prompts, making it more efficient than traditional LLMs 1

   1

   .

Technical Capabilities and Comparisons

While DeepSeek-V3 is a general-purpose model, it has shown promising capabilities in specific areas:

1. Reasoning and Math Skills: The model can solve some math problems and generate code, although it's not specifically optimized for reasoning like its counterpart, DeepSeek-R1 1

   1

   .

2. Competitive Performance: Early testing indicates that the updated V3 model performs better than comparable models like ChatGPT's o3-mini, according to AI entrepreneur Paul Gauthier 2

   2

   .

3. Hardware Compatibility: Awni Hannun, a research scientist at Apple Inc.'s machine learning research group, successfully ran the new DeepSeek-V3 on a high-end Mac Studio, generating output at about 20 tokens per second 1

   1

   .

Related Stories

DeepSeek Unveils Updated R1 AI Model, Challenging Industry Giants

DeepSeek V3: Open-Source AI Model Challenges Industry Giants with Impressive Performance

DeepSeek V3 Upgrade Challenges AI Giants with Open-Source Efficiency

DeepSeek V3.1: A New Contender in the US-China AI Race

Impact on the AI Landscape

The release of the improved DeepSeek-V3 model has broader implications for the AI industry:

1. Open-Source Advancement: By releasing under the MIT License, DeepSeek is contributing to the democratization of AI technology, potentially accelerating innovation in the field 1

   1

   2

   2

   .

2. Chinese AI Capabilities: The update follows the success of DeepSeek's R1 model, which had previously demonstrated China's growing prowess in AI development 2

   2

   .

3. Industry Competition: DeepSeek's advancements have spurred increased activity among Chinese tech giants, with companies like Baidu, Bytedance, Alibaba, and Tencent releasing new AI models to capitalize on the momentum 2

   2

   .

Training and Efficiency

The original DeepSeek-V3 model was trained on a dataset of 14.8 trillion tokens, using approximately 2.8 million graphics card hours – significantly less than what is typically required for frontier LLMs. To enhance output quality, DeepSeek engineers fine-tuned the model using prompt responses from DeepSeek-R1 1

1

As the AI landscape continues to evolve rapidly, DeepSeek's latest release represents a significant step forward in making powerful language models more accessible and efficient for developers and researchers worldwide.