'DeepSeek R1-0528' Introduced As An Open Source Rival To OpenAI o3 And Google Gemini 2.5 Pro

China, vast in both geography and population, has long maintained tight control over its digital borders.

This strategic insulation from foreign influence has fostered an environment where domestic companies thrive, largely free from external competition. With strict regulations governing internet access and the approval of foreign technologies, the country effectively reserves its enormous market—home to over a billion internet users—for homegrown solutions.

This approach has given Chinese tech companies a significant head start in shaping user experiences on their own terms.

As artificial intelligence becomes the global battleground for innovation, sparked in part by the West’s enthusiasm and OpenAI’s launch of ChatGPT, China has not stood idle. A growing number of domestic firms, ranging from industry titans to ambitious startups, are stepping up to stake their claim in this high-stakes arena.

Among these challengers is DeepSeek—a relatively lesser-known name, yet one that is making bold strides.

What truly sets the company apart is its bold introduction of DeepSeek-R1. Developed using comparatively modest hardware and with a more cost-effective training process, the model still manages to rival both OpenAI o1 and other leading contenders. Perhaps even more striking is that it operates with fewer safeguards, allowing it to explore reasoning and expression with notably fewer restrictions.

Now, DeepSeek is updating the R1, making it an even formidable foe of the West.

DeepSeek-R1 stands out for its exceptional reasoning abilities.

Unlike traditional models that rely heavily on supervised fine-tuning, The R1 employs a unique reinforcement learning (RL) training methodology to develop human-like reasoning skills. This approach enables the model to exhibit behaviors such as self-correction, reflection, and the ability to reevaluate its reasoning paths—often referred to as "aha moments."

While the update is kind of minor, but it packs a huge increase in performance.

In fact, the update brings DeepSeek’s free and open model near parity in reasoning capabilities with proprietary paid models such as OpenAI’s o3 and Google Gemini 2.5 Pro.

It also excels in coding, outperforming xAI's Grok 3 on various benchmarks.

This update is geared toward delivering enhanced performance on complex reasoning challenges spanning mathematics, science, business, and programming, while also offering advanced features tailored for developers and researchers.

DeepSeek attributes this improvement to increased computational power and algorithmic optimizations in the post-training phase.

The results speak for themselves.

On the AIME 2025 benchmark, the model’s accuracy jumped from 70% to 87.5%, with average reasoning depth expanding to 23,000 tokens per query—nearly double the previous iteration. Coding performance also advanced, with accuracy on LiveCodeBench increasing from 63.5% to 73.3%, while performance on the notoriously difficult “Humanity’s Last Exam” more than doubled, reaching 17.7%.

What's more, DeepSeek-R1-0528 isn't just about power—it's also built for usability. The update introduces support for JSON output, function calling capabilities, streamlined system prompts, reduced hallucination rates, and a refined front-end user experience for smoother interaction.

To further support the community, the open-source model weights are readily available on Hugging Face AI code sharing platform, complete with comprehensive documentation for those who prefer local deployments or integration via the DeepSeek API.

Existing users of the DeepSeek API can look forward to a seamless transition, as model inferences are automatically updated to R1-0528 at no extra charge.

To address varying computational needs, DeepSeek also introduced DeepSeek-R1-0528-Qwen3-8B, a distilled version that maintains top-tier reasoning in a much smaller footprint.

This model beats Qwen3-8B by 10% on key tasks and rivals the capabilities of significantly larger models like Qwen3-235B-thinking—all while being runnable on a single 16GB GPU. Quantized versions can even run on more modest 8–12GB GPUs, like the RTX 3060.