Why it matters: Techcrunch reports that DeepSeek has unveiled V3, its most advanced AI model to date, achieving benchmark scores that rival proprietary models while offering significantly lower costs. The breakthrough demonstrates growing competition in AI development from open-source alternatives.
The Big Picture: DeepSeek V3 represents major technical advances:
- 671 billion total parameters (The Decoder)
- Processes 60 tokens per second
- Trained on 14.8 trillion tokens
- Cost $5.576 million to develop
Benchmark Performance: The model sets new standards in key tests:
- 90.2% on MATH 500 benchmark (The AI Track)
- 88.5 on MMLU
- 75.9 on MMLU-Pro
- Strong coding performance on Codeforces and SWE
Cost Innovation: DeepSeek challenges industry pricing models:
- $0.27 per million input tokens
- $1.10 per million output tokens
- Cache hits reduced to $0.07 per million tokens
- Current V2 rates maintained until February 8th
Looking Forward: DeepSeek aims to “break through the architectural limitations of Transformer” while pursuing artificial general intelligence through incremental improvements. The company achieved its results using just 2,000 GPUs, compared to the 100,000 typically used by major competitors.
