TL;DR
DeepSeek has released two powerful AI models—V3.2 and V3.2-Speciale—that match or exceed GPT-5 and Gemini-3.0-Pro on key benchmarks. Both 685-billion-parameter models are freely available under an MIT licence, with inference costs reduced by 70% through a novel sparse attention architecture.
Chinese Challenger Disrupts Frontier AI Economics
Hangzhou-based DeepSeek has launched models that could fundamentally reshape AI competition. The standard V3.2 serves as an everyday reasoning assistant, while the Speciale variant achieved gold-medal performance in four elite international competitions: the 2025 International Mathematical Olympiad, International Olympiad in Informatics, ICPC World Finals, and China Mathematical Olympiad.
The benchmark results are striking. On AIME 2025, the prestigious American mathematics competition, DeepSeek-V3.2-Speciale achieved a 96.0% pass rate compared to GPT-5-High’s 94.6%. On coding benchmarks, DeepSeek resolved 73.1% of real-world software bugs on SWE-Verified, competitive with GPT-5-High at 74.9%.
Sparse Attention Slashes Computing Costs
At the heart of the release lies DeepSeek Sparse Attention (DSA), an architectural innovation that dramatically reduces computational burden. Traditional attention mechanisms scale poorly with input length—processing a document twice as long typically requires four times the computation. DeepSeek’s approach uses a “lightning indexer” to identify only the most relevant context portions.
The cost implications are significant: processing 128,000 tokens—roughly equivalent to a 300-page book—now costs approximately $0.70 per million tokens, compared to $2.40 for the previous model. That represents a 70% reduction in inference costs.
Open-Source Strategy Challenges Business Models
Unlike OpenAI and Anthropic, DeepSeek has released both models under the MIT licence. Any developer can download, modify, and deploy the 685-billion-parameter models without restriction. For enterprise customers, this offers frontier performance at dramatically lower cost—though data residency concerns may limit adoption given DeepSeek’s Chinese origins.
Looking Forward
Regulatory walls are rising against DeepSeek in Europe and America, with Germany declaring user data transfers to China “unlawful” and Italy ordering app blocks. Nevertheless, the release demonstrates that export controls alone cannot halt Chinese AI progress, suggesting the competitive landscape may shift faster than many anticipated.
Source: VentureBeat