DeepSeek, a Chinese artificial intelligence startup, has significantly upgraded its open-source reasoning model with the release of R1-0528, bringing its performance nearly on par with industry leaders like OpenAI's o3 and Google's Gemini 2.5 Pro.
The upgrade represents a major advancement in DeepSeek's reasoning capabilities through increased computational resources and algorithmic optimization. In benchmark tests, the model showed remarkable improvements in mathematical reasoning, with accuracy on the AIME 2025 math test increasing from 70% to 87.5%. This improvement stems from enhanced reasoning depth, with the model now using an average of 23,000 tokens per question compared to the previous 12,000.
Beyond mathematical improvements, R1-0528 offers several enhanced features that make it more developer-friendly. The model now supports system prompts, function calling, and JSON output, making it easier to integrate into applications. It also demonstrates a reduced hallucination rate—reportedly down by 45-50% in rewriting and summarizing tasks—and provides better support for vibe coding, where developers use natural language to generate code.
In a notable development for resource-constrained users, DeepSeek also released a distilled version called DeepSeek-R1-0528-Qwen3-8B. Created by fine-tuning Alibaba's Qwen3 8B model with the reasoning patterns from the full R1-0528, this smaller model can run on a single consumer-grade GPU while still outperforming Google's Gemini 2.5 Flash on certain math benchmarks and nearly matching Microsoft's Phi 4 reasoning plus model.
Like its predecessor, R1-0528 is available under the permissive MIT License, allowing for commercial use and customization. Major cloud providers including Amazon Web Services and Microsoft Azure now offer DeepSeek's models to their clients, though they isolate them from Chinese servers to ensure data privacy.
This release further cements DeepSeek and Meta as leaders in the open-source AI space, providing powerful alternatives to proprietary models from OpenAI, Google, Microsoft, and Anthropic at a fraction of the computational cost.