China’s DeepSeek Drops a Free GPT 5.1 Rival for Completely Free

Chinese startup DeepSeek has released two new large AI models, DeepSeek-V3.2 and DeepSeek-V3.2-Special, entering an already intense international competition. DeepSeek-V3.2 is designed for everyday reasoning, while the Speciale variant targets advanced math and coding tasks. Both models were published under an open-source license, meaning they are free to run and use for anyone.

Open-Source Release Stands Out

The significance of the release extends beyond model performance. While US firms such as OpenAI and Google deploy powerful and often costly models through private APIs and rely on controlled testing for their top systems, DeepSeek has chosen an open approach.

DeepSeek-V3.2 reportedly matches or surpasses GPT-5 and Gemini 3 Pro in long-form reasoning, tool use, and complex problem solving, including results in competitions such as the International Mathematical Olympiad and the ICPC World Finals.

The Speciale model scored 99.2% on the Harvard-MIT Math Tournament, reached 73% in software bug-fixing tasks, and achieved gold-medal outcomes on multiple international benchmarks without internet access or external tools.

New Architecture Cuts Compute Costs

The models rely on an architectural method known as DeepSeek Sparse Attention, or DSA. Traditional transformer models expand in cost as context length increases because they compare every word with every other word. DSA reduces costs by focusing only on the most relevant portions of the input, cutting long-document expenses by up to 70%.

This reduction matters for deployment. Many of today’s frontier models sit behind paywalls or throttled access, but DeepSeek’s new systems and their 128,000-token context windows are free to download and modify. This allows individual developers or small teams to work with models that previously required significant cloud budgets or research labs.

Improvements in Tool Use

DeepSeek also highlights progress in tool-use reasoning. Many AI agents fail to manage multiple tools because each action resets internal reasoning. DeepSeek addressed this by preserving memory across tools. The company trained the models on more than 85,000 complex synthetic instructions involving real web browsers and coding environments.

These changes support tasks that require linked decision-making. Beyond simple text generation, the models can plan multi-day trips with strict budgets, test code, and check exchange rates while managing interdependent constraints.

Open Licensing Sparks Regulatory Concern

DeepSeek’s decision to adopt the MIT open-source license may have broader consequences. The license allows anyone to copy, modify, or commercialize the models, pushing against industry efforts to restrict access to model weights over safety and misuse concerns.

However, openness does not guarantee transparency. German regulators have attempted to block DeepSeek over data-transfer issues. Italy banned the app earlier this year, and US lawmakers want it removed from government devices. DeepSeek’s status as a Chinese company continues to influence regulatory discussions.

Wider Access Expected in December

For now, the Speciale variant is available only through a temporary API. DeepSeek plans to merge it into the broader V3.2 release by mid-December, making it publicly accessible. After years shaped by ChatGPT’s introduction to mainstream AI, DeepSeek’s move signals a shift toward a competition defined by access, cost, and control.

The post China’s DeepSeek Drops a Free GPT 5.1 Rival for Completely Free appeared first on ProPakistani.

Chinese startup DeepSeek has released two new large AI models, DeepSeek-V3.2 and DeepSeek-V3.2-Special, entering an already intense international competition. DeepSeek-V3.2… Read More

The post China’s DeepSeek Drops a Free GPT 5.1 Rival for Completely Free appeared first on ProPakistani.