DeepSeek launches world's best non-thinking LLM—How it compares against ChatGPT?
What's the story
Chinese AI start-up, DeepSeek, has launched an upgraded version of its V3 large language model.
Dubbed DeepSeek-V3-0324, it offers improved coding and math problem-solving capabilities.
It is being touted as the world's best non-reasoning AI model, demonstrating superior performance across key benchmarks compared to leading AI models like OpenAI's GPT-4.5 and Athropic's Claude 3.7 Sonnet.
The new version comes with "enhanced reasoning capabilities," optimized front-end web development skills as well as improved Chinese writing proficiency.
Model comparison
DeepSeek-V3-0324 outperforms GPT-4.5 across multiple key evaluation metrics
In the MATH-500 benchmark, which tests advanced mathematical problem-solving skills, DeepSeek-V3-0324 achieves a remarkable 94% accuracy, outperforming GPT-4.5's 90.7%, showcasing its superior numerical reasoning abilities.
In the AIME 2024 test, a highly challenging math competition benchmark, DeepSeek-V3-0324 dominates with a 59.4% success rate, nearly double GPT-4.5's 36%.
In the field of coding, the LiveCodeBench test, which evaluates programming ability and code generation accuracy, also favors DeepSeek-V3-0324, which achieves a 49.2% pass rate, compared to GPT-4.5's 44%.
Technical specifications
DeepSeek-V3-0324 is also cheaper to run
DeepSeek is significantly more cost-effective than GPT-4.5, with token processing priced at $0.27 and $1.10 per million tokens, compared to GPT-4.5's $75 and $150.
It also runs five times faster, processing 60 tokens per second versus GPT-4.5's 12.
Despite its efficiency, DeepSeek is smaller, using a 685B Mixture of Experts (MoE) model, while GPT-4.5 is estimated at two trillion parameters.
Moreover, DeepSeek is free to distribute under the permissive MIT license and is also open source, allowing wider adoption.
Twitter Post
Take a look at DeepSeek's performance chart
BREAKING DeepSeek has #1 best non-thinking LLM.
— Deedy (@deedydas) March 25, 2025
— Better (beats or ties GPT4.5, etc)
— Cheaper, by 100-200x ($0.27/1.10 vs $75/$150 for 1M input/output toks)
— Faster, by 5x (60 tok/s vs ~12 tok/s)
— Smaller (685B MoE vs 2T??)
— Free to distribute (MIT license)
— Open source pic.twitter.com/PVc9nof0o0