DeepSeek, a Chinese AI research lab backed by High-Flyer Capital Management has released DeepSeek-V3, the latest version of their frontier model. The Mixture-of-Experts model features a total of 671B total parameters, with 37B activated for each token. The model has been trained on 14.8 trillion tokens. DeepSeek has released the model on GitHub and a detailed technical paper outlining its capabilities.
DeepSeek AI also released the benchmark scores, and it outperformed Meta’s flagship Llama 3.1 405B parameter model, among many other closed-source models. It is also three times faster than its predecessor, the DeepSeek V2.
“Comprehensive evaluations demonstrate that DeepSeek-V3 has emerged as the strongest open-source model currently available and achieves performance comparable to leading closed-source models like GPT-4o and Claude-3.5-Sonnet,” read the technical paper.
Moreover, DeepSeek also mentioned that it has distilled its reasoning capabilities from the DeepSeek R1 series of models.
“Our pipeline elegantly incorporates the verification and reflection patterns of R1 into DeepSeek-V3 and notably improves its reasoning performance,” read the paper.
Moreover, the API pricing will remain the same as DeepSeek V2 until February 8, 2025. Further on, it will cost $0.27/million tokens during input and $1.10/million tokens during output.
Meaning this also makes it one of the cheapest models out there.
While it may not be a fair comparison, how does the model fare with OpenAI’s o1? While o1 scored a 76% score on the GPQA Diamond (PhD-Level Science Questions) benchmark, DeepSeek does lag behind with a 59.1% score. The full version of o1 beats DeepSeek on multiple benchmarks. However, DeepSeek-V3 does outperform the coveted Claude 3.5 Sonnet across multiple benchmarks.
That said, DeepSeek has been taking major strides in the open-source AI ecosystem over the last few months. Only a few weeks ago did the company launch the V2.5-1210, the final model in its V2 series. The model is accessible on chat.deepseek.com. Users can toggle the Internet Search feature on the website for real-time responses or integrate the model via Hugging Face.
Models from the east are giving the ones from the west a run for their money, and DeepSeek isn’t the only one. Alibaba’s Qwen 2.5 on the other hand, offered performance parity with many leading models. The Qwen2.5-Coder series excels in code generation, matching the capabilities of GPT-4o on benchmarks like EvalPlus, LiveCodeBench, and BigCodeBench.
The post DeepSeek-V3 is Now The Best Open Source AI Model appeared first on Analytics India Magazine.
