DeepSeek-R1 from China Surpasses US AI Giants, Creating Shockwaves in Silicon Valley
Silicon Valley was abuzz this week as DeepSeek, a Chinese AI company, unveiled its latest model, the R1, which has outperformed leading US AI giants like OpenAI, Meta, and Anthropic in independent benchmarks. The tech community was already taken aback by DeepSeek's v3 model, which outpaced Meta's Llama 4 in previous tests. However, the real surprise came on Jan. 20 with the release of DeepSeek-R1, showcasing remarkable improvements in model inference capabilities despite having limited labelled data.
DeepSeek-R1 has proven to be a formidable competitor, matching OpenAI's GPT-4 in various tasks such as mathematics, coding, and natural language reasoning. What's particularly intriguing is the attention drawn to DeepSeek-R1's training budget. The company's published API pricing reveals a significantly lower cost structure compared to its US counterparts. For instance, it charges RMB 1 ($0.14) per million input tokens for cache hits, RMB 4 ($0.55) for cache misses, and RMB 16 ($2.21) for million output tokens. This pricing model is approximately one-thirtieth of the operational expenses incurred by OpenAI for GPT-4.
Yan LeCun, Meta's chief AI scientist, weighed in on DeepSeek's success, highlighting how open-source models like DeepSeek are now surpassing proprietary ones.
DeepSeek, a Chinese AI firm, has launched the R1 model, surpassing top US AI companies in benchmarks.
DeepSeek-R1 performs on par with OpenAI's GPT-4 in tasks like mathematics, coding, and natural language reasoning.
The pricing structure of DeepSeek-R1's training budget is significantly lower than that of its US counterparts.
Source: TECHNODE