The beginning of Chinese artificial intelligence Dipsec On Sunday, it released two powerful new AI models that the company claims surpass or surpass 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.
Launched by the Hangzhou-based company Depsec-V 3.2designed as an everyday reasoning aid, along with DEPESEC-V3.2-Special, in a high-powered variant, achieved gold medal performance in four elite international competitions: the 2025 International Mathematical Olympiad, the International Olympiad in Informatics, the ICPC World Finals, and the China Mathematical Olympiad.
The release has profound implications for American technology leadership. Depsec has once again demonstrated that it can develop frontier AI systems despite US export controls. Ban China’s access to advanced NVIDIA chips -and it has done so by making its models freely available under the open-source MIT license.
"People thought that Dipsec gave a temporary breakthrough but we came back big," It is written Chen Fengwhich on X (formerly Twitter), identified itself as a contributor to the project. The release sparked a swift reaction online, with one user declaring: "Rest in peace, Chetgupt."
How Deepsec’s Discrete Attention Advances Reduce Computing Costs
lies at the heart of the new release Depesic sparse attentionor DSA – a novel architectural innovation that dramatically reduces the computational burden of running AI models on long documents and complex tasks.
Unlike traditional AI attention mechanisms, the underlying technology allows the language model to understand context, as the length of the input increases. Processing a document twice as long usually requires four times the computation. Dipsec’s approach breaks down the barrier of what is called a company "Power index" which indicates only the most relevant parts of the context for each question, ignoring the rest.
According to Deepak Technical ReportDSA cuts costs by nearly half compared to previous models when processing long sequences. Architecture "substantially reduces the computational complexity while preserving the performance of the model," The report said.
Processing 128,000 tokens—the equivalent of about a 300-page book—now costs about $0.70 per million tokens to decrypt, compared to $2.40 for the previous 40. V3.1-terminus model. That represents a 70% reduction in mitigation costs.
685 billion parameter models support context windows of 128,000 tokens, making them suitable for analyzing long documents, codebases and research papers. Of Deepacek Technical report Note that independent evaluations on the Long Context benchmark show V3.2 performing on par with or better than its predecessor. "Despite incorporating a sparse attention mechanism."
Benchmark results that put Deepsec in the same league as GPT5
Claims of parity with America’s leading AI systems rest on extensive testing in math, coding, and reasoning tasks. And this number is amazing.
But AIME 2025a prestigious American math competition, Depsec-V 3.2 Spacial Achieved a 96.0% pass rate, compared to 94.6% for GPT-5-Hi and 95.0% for Gemini-3.0-Pro. But Harvard-MIT Mathematics Tournamentthe Spacial variant scored 99.2%, beating Gemini’s 97.5%.
Quality V3.2 modeloptimized for everyday use, scored 93.1% on AIME and 92.5% on HMMT – marginally below the frontier models but achieved with considerably less computational resources.
Most surprising are the results of the competition. Depsec-V 3.2 Spacial Scored 35 runs out of 42 points on it 2025 International Mathematical Olympiadachieving gold medal status. But International Olympiad in Informaticshe scored 492 out of 600 points – also gold, ranking 10th overall. The model solved 10 out of 12 problems ICPC World Finalsto hold another.
These results came without internet access or tools during testing. It is stated in the Dipsec report that "Strictly checking on competition time and effort limits."
On the coding benchmark, Depsec-V 3.2 Solved 73.1% of real-world software bugs SWE-certified74.9% competitive with GPT-5-HI. But Terminal Bench 2.0measuring complex coding workflows, Dipsec scored 46.4%—above GPT5-High’s 35.2%.
Company acknowledges limitations. "Token performance remains a challenge," The technical report states that it notes the dipsec "A long generation path is usually required" To match the output quality of Gemini-3.0-Pro.
Why teaching AI to think using tools changes everything
Beyond crude reasoning, Depsec-V 3.2 Introduces "Thinking in tool use" – Ability to simultaneously execute code, search the web, and reason through problems while 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 restart the reasoning from scratch. DeepSec’s architecture provides logic across multiple tool calls, enabling fluid multi-step problem solving.
To train this capability, the company has built a massive simulation data pipeline that generates more than 1,800 distinct work environments and 85,000 complex instructions. They require multi-day trip planning with budget constraints, software bug fixes in eight programming languages, and dozens of searches like web-based research.
An example is described in the technical report: planning a three-day trip from Hangzhou with constraints on hotel prices, restaurant ratings, and attraction costs that vary based on accommodation choices. There are such works "Difficult to solve but easy to verify," Making them ideal for training AI agents.
Dipsec The training used real-world tools—real web search APIs, coding environments, and Jupiter notebooks—when generating synthetic signals to ensure diversity. The result is a model that generalizes unseen tools and environments, a critical capability for real-world deployments.
Deepsec’s open-source gambit could upend the AI ​​industry’s business model
Unlike Openei and Anthropic, which keep their most powerful models as proprietary assets, Dipsec has released both v3.2 And v3.2-speciale The most permissive open source framework available—under the MIT License.
Any developer, researcher, or company can download, modify, and deploy the 685-billion-parameter model without restriction. There are complete model weights, training code, and documentation Available in face huggingthe leading platform for AI model sharing.
The strategic implications are significant. By making Frontier’s enablement model freely available, Deepsec undermined competitors charging premium API prices. Notes on the Hugging Face model card provided by Dipsec with Python scripts and test cases "Demonstrating how to encode messages in an OpenAI-compatible format" – Migrating directly from competing services.
For enterprise users, the value proposition is compelling: frontier performance at a dramatically lower cost, with flexibility in deployment. But data-housing concerns and regulatory uncertainty could limit adoption in sensitive applications — especially given Deepsec’s Chinese origins.
Regulatory walls are mounting against Dipsec in Europe and the US
Dipsec’s global expansion faces growing resistance. In June, Berlin’s data protection commissioner Mike Kemp announced that Deepsec’s German user data was to be transferred to China. "illegal" Under EU rules, Apple and Google are asked to consider blocking the app.
The German authority expressed this concern "Chinese authorities have broad rights to access personal data within the sphere of influence of Chinese companies." Italy ordered Deep Sak Block his app US lawmakers moved in February Ban the service from government instruments citing national security concerns.
Questions also remain about US export controls designed to limit China’s AI capabilities. In August, Deepak hinted that China would soon "The next generation" Chips were made domestically to support their own models. The company indicated its system will work with Chinese-made chips Huawei And Cambrican Without additional setup.
Depict’s original V3 model was reportedly trained around 2,000 years ago. Nvidia H800 chips – Hardware as limited for export to China. The company hasn’t revealed what powered V3.2’s training, but its continued progress suggests that export controls alone can’t stop Chinese AI progress.
What Deepak’s release means for the future of AI competition
The release comes at a critical moment. After years of massive investment, some analysts question whether an AI bubble is forming. Deepsec’s ability to match US frontier models at a fraction of the cost challenges challenges assumptions that AI leadership requires huge capital expenditures.
of the company Technical report reveals that post-training investment is now more than 10% of pre-training expenditure. But Depesek acknowledged the gap: "The extent of global knowledge in Depsec-v3.2 still lags behind well-known proprietary models," The report said. The company plans to solve this by scaling pre-training compute.
Depsec-V 3.2 Spacial Available through a temporary API until December 15th, when its capabilities will be integrated into the standard release. The Spacial variant is designed exclusively for deep reasoning and does not support tool calling—a limitation the standard model addresses.
For now, the AI ​​race between the United States and China has entered a new phase. The release of Deepsec demonstrates that open-source models can achieve frontier performance, that performance innovations can dramatically reduce costs, and that the most powerful AI systems can be freely available to anyone with an Internet connection.
As one commenter on X observed: "Depesek just casually breaking the historical standards set by Gemini is bonkers."
The question is no longer whether Chinese AI can compete with Silicon Valley. It’s whether American companies can maintain their edge when their Chinese rivals give away comparable technology for free.