Alibaba open-sources Qwen2.5-VL-32B: small model, big power

• • DeepSeek suddenly released a new open source model V3-0324, which is an enhanced version of V3 released in December last year. It uses the MIT license and is available to everyone.
• • The community feedback was overwhelming. The code writing and logical reasoning capabilities of V3-0324 were significantly improved, but the conversation style seemed to be more "technical".
• • Industry insiders speculate that this model is likely to be a “pre-release” of DeepSeek’s upcoming R2 inference model.
• • The model is now available for free on DeepSeek’s official website and Hugging Face, go and experience it now!

DeepSeek V3-0324: An open source rising star

On March 24, 2024, DeepSeek, a "dark horse" company in the field of domestic AI, once again surprised the open source community by quietly releasing its latest masterpiece, DeepSeek V3-0324. The name of this model is concise and clear, and "0324" directly indicates its birth date.

DeepSeek V3-0324 is an upgraded version of DeepSeek V3 released in December last year. V3-0324 adopts the Mixed Expert Model (MoE) architecture. More importantly, it fully complies with the MIT license agreement, which means that developers and researchers around the world can use, modify and even distribute this model for free without any restrictions!

Zero threshold experience: DeepSeek V3-0324 free trial guide

Want to experience the charm of DeepSeek V3-0324 for yourself? There are two free ways for you to choose:

1. 1.  DeepSeek official website : One step directly to DeepSeek's online chat platform (https://chat.deepseek.com/), chat with V3-0324 directly, no configuration required, experience it immediately!
2. 2.  Hugging Face Model Library : If you are an AI developer or researcher who likes to build your own environment, you can find the complete model weights and related resources of V3-0324 in Hugging Face (https://huggingface.co/deepseek-ai/DeepSeek-V3-0324), download it freely, and deploy it at will!

Community Explosion: DeepSeek V3-0324 First Experience Report

Although DeepSeek has not yet released a detailed performance evaluation report for V3-0324, community users who have tried it can’t wait to share their initial experiences:

• •  Big upgrade in coding ability : Many developers exclaimed that V3-0324's performance in coding is amazing, and it has made a qualitative leap compared to the previous DeepSeek model. "It can write code quickly and accurately, which is really good news for programmers!" a user said excitedly.
• •  Improved reasoning ability : A user tested V3-0324 with complex logical reasoning tasks such as "a bouncing ball in a rotating shape" and the results were surprising. "It can accurately understand the meaning of the question and give a reasonable answer. It has strong logical thinking ability."
• •  Technical dialogue style : Interestingly, many users found that the dialogue style of V3-0324 seemed to become more "mechanical" and more "technical expert" compared to the previous V3, and lacked some of the "human touch" of V3. One user described it this way: "It feels more like a knowledgeable engineer than a chatbot."

DeepSeek V3-0324: Is this a trailer for R2?

The sudden release of V3-0324 has led AI insiders to speculate that this is likely a "warm-up" for DeepSeek's upcoming R2 inference model. After all, DeepSeek released V3 at the end of last year, and then launched the powerful R1 at the beginning of this year.

At this pace, V3-0324 is likely to be the "advanced version" of R2. According to various sources, R2 is expected to be officially unveiled in April or May 2024. Let us wait and see!

DeepSeek V3-0324: A new chapter of open source, promising future

• • Potential impact on the open source large model community
• • Continuation of the open source spirit: DeepSeek has unswervingly followed the open source route, contributing valuable resources to the global AI community and promoting the sharing and advancement of technology.
• • Inspiration for other research institutions: DeepSeek’s innovations in model architecture, training methods, etc. may provide new inspiration and directions for other AI research teams.
• • Thoughts triggered by changes in model style
• • Is the “mechanized” style the future trend? V3-0324 The change in conversation style forces us to think: What is the development direction of AI models? Is it more like a “human” or more like a “tool”?
• • The trade-off between user experience and model performance: While pursuing extreme performance, should we pay more attention to the user experience of interacting with AI? After all, AI is ultimately for the service of humans.