December 26, 2024:
DeepSeek-V3 Sets New Benchmark in Open AI - Chinese startup DeepSeek launched DeepSeek-V3, an ultra-large open-source AI model with 671B parameters, outperforming top models like Llama and Qwen. It utilizes a mixture-of-experts architecture and achieves superior efficiency with innovations such as auxiliary loss-free load-balancing and multi-token prediction. Trained at a fraction of the usual cost, it excels in Chinese and math benchmarks, challenging closed models like GPT-4o.
DeepSeek-V3 marks a significant step toward closing the gap between open and closed-source AI, potentially reshaping the AI landscape by offering enterprises competitive alternatives.
