DeepSeek just dropped two insanely powerful AI models that rival GPT-5 and they're totally free
Chinese artificial intelligence startup Deep Search on Sunday released two powerful new AI models that the company claims will match or even exceed OpenAI’s capabilities GPT-5 and Google’s Gemini-3.0-Pro – a development that could reshape the competitive landscape between American tech giants and their Chinese challengers.
The Hangzhou-based company was launched DeepSeek-V3.2designed as an assistant for everyday reasoning, alongside DeepSeek-V3.2-Speciale, a powerful variant that achieved gold medal performances in four elite international competitions: the 2025 International Mathematical Olympiad, the International Olympiad in Computer Science, the ICPC World Finals and the Chinese Mathematical Olympiad.
The release has profound implications for America’s technology leadership. DeepSeek has once again demonstrated its ability to produce groundbreaking AI systems despite US export controls restrict China’s access to advanced Nvidia chips – and the company has done so by making its models freely available under an open-source MIT license.
“People thought DeepSeek was a one-time breakthrough, but we came back much bigger,” wrote Chen Fangwho identified himself as a contributor to the project, on X (formerly Twitter). The release provoked swift reactions online, with one user stating: “Rest in peace, ChatGPT.”
How DeepSeek’s sparse attention breakthrough reduces computing costs
The core of the new release lies Deep Seek scarce attentionor DSA – a new architectural innovation that dramatically reduces the computational burden of running AI models on long documents and complex tasks.
Traditional AI attention mechanisms, the core technology that allows language models to understand context, scale poorly as input length increases. Processing a document for twice as long usually requires four times as many calculations. DeepSeek’s approach breaks this limitation by using what the company calls a “lightning indexer,” which identifies only the most relevant parts of the context for each search query and ignores the rest.
According to The DeepSeek technical reportDSA reduces inference costs by about half compared to previous models when processing long sequences. The architecture “significantly reduces computational complexity while maintaining model performance,” the report said.
Processing 128,000 tokens – roughly equivalent to a 300-page book – now costs around $0.70 per million tokens to decrypt, compared to $2.40 for the previous V3.1 Terminus model. This means a 70% reduction in inference costs.
The 685 billion parameter models support context windows of 128,000 tokens, making them suitable for analyzing long documents, codebases, and research papers. DeepSeeks technical report notes that independent evaluations of long-context benchmarks show that V3.2 performs as well as or better than its predecessor, “despite including a sparse attention mechanism.”
The benchmark results that put DeepSeek in the same league as GPT-5
DeepSeek’s claims of parity with America’s leading AI systems rely on extensive testing in math, coding and reasoning – and the numbers are striking.
On AIME 2025a prestigious American mathematics competition, DeepSeek-V3.2-Special achieved a pass rate of 96.0%, compared to 94.6% for GPT-5-High and 95.0% for Gemini-3.0-Pro. On the Harvard-MIT Math Tournamentthe Special variant scored 99.2%, surpassing Gemini’s 97.5%.
The standard V3.2 modeloptimized for everyday use, it scored 93.1% on AIME and 92.5% on HMMT – slightly below the borderline models, but achieved with significantly less computing resources.
The most striking are the competition results. DeepSeek-V3.2-Special scored 35 out of 42 points on the International Mathematical Olympiad 2025that deserves gold medal status. At the International Olympiad in Computer Scienceit scored 492 out of 600 points – also gold, and ranked 10th overall. The model solved 10 of the 12 problems ICPC World Finalscame second.
These results came during testing without internet access or tools. DeepSeek’s report states that “testing strictly adheres to the time and attempt limits of the competition.”
As for coding benchmarks, DeepSeek-V3.2 fixed 73.1% of real software bugs SWE verifiedcompetitive with GPT-5-High at 74.9%. On Terminal bank 2.0When measuring complex coding workflows, DeepSeek scored 46.4%, well above GPT-5-High’s 35.2%.
The company recognizes limitations. “Token efficiency remains a challenge,” the technical report states, noting that DeepSeek “typically requires longer generation paths” to match the output quality of Gemini-3.0-Pro.
Why AI learning to think while using tools changes everything
Besides a rough reasoning, DeepSeek-V3.2 introduces ‘thinking in tools’: the ability to reason through problems while simultaneously executing code, searching the web, and manipulating files.
Previous AI models suffered from a frustrating limitation: every time they called an external tool, they lost their train of thought and had to reason all over again. DeepSeek’s architecture preserves the reasoning trail across multiple tool calls, enabling fluid multi-step problem solving.
To train this capability, the company built a massive synthetic data pipeline that generates more than 1,800 different task environments and 85,000 complex instructions. These included challenges such as planning multi-day trips on limited budgets, troubleshooting software bugs in eight programming languages, and web-based research that required dozens of searches.
The technical report describes one example: planning a three-day trip from Hangzhou with restrictions on hotel prices, restaurant ratings and attraction costs that vary based on accommodation choices. Such tasks are “hard to solve but easy to verify,” making them ideal for training AI agents.
Deep Search used real-world tools (real web search APIs, coding environments, and Jupyter notebooks) during the training while generating synthetic prompts to ensure diversity. The result is a model that generalizes to invisible tools and environments, a crucial capability for real-world implementation.
DeepSeek’s open source strategy could shake up the AI industry’s business model
Unlike OpenAI and Anthropic, which guard their most powerful models as proprietary assets, DeepSeek has released both V3.2 And V3.2-Special under the MIT license – one of the most permissive open source frameworks available.
Any developer, researcher or company can download, modify and deploy the models with 685 billion parameters without restrictions. Full model weights, training code and documentation are provided available on Cuddle Facethe leading platform for sharing AI models.
The strategic implications are significant. By making Frontier-compatible models freely available, DeepSeek undercuts competitors who charge premium API prices. The Hugging Face model card notes that DeepSeek has provided Python scripts and test cases “that demonstrate how to encode messages in an OpenAI-compatible format” – making migration from competing services easy.
For business customers, the value proposition is compelling: groundbreaking performance at significantly lower costs, with flexibility in implementation. But data location concerns and regulatory uncertainty could limit adoption in sensitive applications – especially given DeepSeek’s Chinese origins.
Regulatory walls rise against DeepSeek in Europe and America
DeepSeek’s global expansion is encountering increasing resistance. In June, Berlin Data Protection Commissioner Meike Kamp stated that DeepSeek’s transfer of German user data to China “unlawful” under EU rules, asking Apple and Google to consider blocking the app.
The German authority expressed concern that “Chinese authorities have extensive access rights to personal data within the sphere of influence of Chinese companies.” Italy has commissioned DeepSeek to do this block the app in February. US lawmakers have switched prohibit the service of government apparatus, citing national security concerns.
Questions also remain about US export controls intended to limit Chinese AI capabilities. In August, DeepSeek hinted that China will soon “next generation“domestically built chips to support its models. The company indicated that its systems work with Chinese-made chips from Huawei And Cambricon without additional settings.
DeepSeek’s original V3 model was reportedly trained on approximately 2,000 older models Nvidia H800 chips – hardware since restricted for Chinese export. The company didn’t reveal what’s driving the V3.2 training, but the continued progress suggests that export controls alone won’t hold back China’s AI progress.
What the release of DeepSeek means for the future of AI competition
The release comes at a crucial time. After years of massive investment, some analysts are wondering if an AI bubble is forming. DeepSeek’s ability to match US frontier models at a fraction of the cost challenges the assumption that AI leadership requires massive capital expenditures.
The company technical report shows that post-training investments now exceed 10% of pre-training costs – a substantial allocation credited with improvements in reasoning. But DeepSeek acknowledges gaps: “The breadth of global knowledge in DeepSeek-V3.2 still lags behind leading proprietary models,” the report states. The company plans to address this by scaling up pre-training computing.
DeepSeek-V3.2-Special will remain available via a temporary API until December 15, when its capabilities will be merged into the standard release. The Special variant is designed exclusively for deep reasoning and does not support tool calling – a limitation that the Standard model addresses.
For now, the AI race between the United States and China has entered a new phase. The release of DeepSeek shows that open source models can deliver groundbreaking performance, that innovations in efficiency can dramatically reduce costs, and that the most powerful AI systems may soon be freely available to anyone with an internet connection.
As one commenter on
The question is no longer whether Chinese AI can compete with Silicon Valley. At issue is whether American companies can maintain their lead if their Chinese rival gives away similar technology for free.
