DeepSeek had stunned the world with the release of DeepSeek R1, but it turns out that R1 was no flash in the pan.
Chinese AI company DeepSeek has released a new version of its V3 model. V3 is the company’s non-reasoning model, which it had first released in December 2024. DeepSeek has now released an updated endpoint of the model, which appears to be doing quite well of many benchmarks.
As per the KCORES LLM Arena, DeepSeek is the second best-performing non-reasoning model in the world behind Claude 3.5 Sonnet. KCORES LLM Arena is a coding benchmark that tests models on four coding tasks. DeepSeek’s new V3 model scored better than Claude 3.7 Sonnet, DeepSeek’s own R1 reasoning model, OpenAI’s o1 model, and Google’s Gemini-2.0-Pro-Experimental model. It was only behind Claude 3.7 Sonnet-Thinking and Claude 3.5 Sonnet on the benchmark.
DeepSeek performed similarly well on the Aider polyglot benchmark. This benchmark asks the LLMs to edit source files to complete 225 difficult coding exercises. On this benchmark, DeepSeek V3’s new version was again the second best non-reasoning LLM behind Claude 3.7 Sonnet. It placed fourth overall, behind Claude 3.7-Thinking, Claude 3.7-No Thinking, and DeepSeek R1. Interestingly, this benchmark also reports the costs it took to solve the problems, and here DeepSeek was by far the cheapest, costing around $1 to solve problem. In comparison, the best model, Claude 3.7-Sonnet cost around $35, while Claude Sonnet 3.7-No Thinking cost around $15. Interestingly, DeepSeek V3 performed better than OpenAI’s o3-mini (medium) model, while also being much cheaper.
These are extremely impressive results, and made even more impressive by the fact that V3 is an open-source model. It’s been released with an MIT license, which is one of the most liberal software licenses going around, which lets users modify and change the model as they see fit. In comparison, some of OpenAI’s models are more expensive, less performant, and not open source. OpenAI might have some aces up its sleeve with newer releases like GPT-5, but it appears that at the moment, DeepSeek is eating the US’s lunch in the AI space.