A naukri.com initiative
VentureBeat
1w
302
Image Credit: VentureBeat
DeepSeek-V3 now runs at 20 tokens per second on Mac Studio, and that’s a nightmare for OpenAI
- Chinese AI startup DeepSeek has released a new large language model, DeepSeek-V3-0324, under an MIT license allowing commercial use.
- The model can run on Apple's Mac Studio with M3 Ultra chip, achieving over 20 tokens per second.
- DeepSeek's launch lacked typical fanfare, with no whitepaper or marketing, but the model has shown improvements over its predecessor.
- DeepSeek-V3-0324 operates with a MoE architecture, activating only 37 billion out of its 685 billion parameters for specific tasks, enhancing efficiency.
- The new model incorporates MLA and MTP technologies, boosting output speed by nearly 80%.
- With a 4-bit quantized version offering reduced storage footprint, it can run on high-end consumer hardware, challenging traditional AI infrastructure.
- Chinese AI companies like DeepSeek opt for open-source licensing, contrasting with Western companies keeping models behind paywalls.
- This strategy enables rapid transformation and AI innovation in China, with tech giants like Baidu, Alibaba, and Tencent also embracing open-source models.
- DeepSeek-R2, an advanced reasoning model, is anticipated to build upon DeepSeek-V3-0324, potentially competing with models like GPT-5 from OpenAI.
- By democratizing access to AI technology through open-source models, DeepSeek is reshaping the future of AI development and adoption globally.
- DeepSeek's approach reflects a broader trend towards making AI more accessible and empowering a wider range of developers and researchers in the field.
Read Full Article
18 Likes
For uninterrupted reading, download the app