DeepSeek has released an updated version of its R1 reasoning model, now available on Hugging Face under the permissive MIT license.
Highlights
- DeepSeek has released R1-0528, a minor but significant upgrade to its flagship R1 model, now hosted on Hugging Face under the MIT license.
- Massive scale: The model boasts 685 billion parameters, positioning it as one of the largest open-source AI models to date—targeted at enterprise and research applications.
- Open and permissive licensing: The MIT license allows unrestricted commercial use, modification, and integration into proprietary products.
- Competitive performance: Ranks just behind OpenAI’s o4-mini and o3 on LiveCodeBench, outperforming Grok-3-mini and Qwen-3 in code generation tasks.
- Innovative training: Utilizes reinforcement learning without supervised fine-tuning, enhancing chain-of-thought reasoning and multi-step capabilities.
- Distilled variants available: Includes Llama and Qwen-based versions for more accessible deployment, with one outperforming OpenAI’s o1-mini in multiple tests.
- Documentation gaps: Despite the release, Hugging Face lacks in-depth documentation, deployment examples, or fine-tuning guidelines.
- Strategic move: DeepSeek aims to position itself as a global open-source AI leader while navigating geopolitical scrutiny from U.S. regulators.
- Broader implication: The release signifies the growing maturity of China’s open-source AI scene and the expanding competitive landscape beyond the West.
- More than a “minor” update: R1-0528 reinforces DeepSeek’s momentum and provides a strong alternative for high-performance, open AI development.
While the company describes the release as a “minor” upgrade, it reflects continued progress in its open-source AI efforts and growing presence in the global AI ecosystem.
The announcement was shared via WeChat and highlights incremental improvements to the R1 model, which has been positioned as a notable open-source alternative to proprietary models from larger U.S.-based organizations such as OpenAI.
Model Details
The updated version, referred to as R1-0528, features a substantial 685 billion parameters, placing it among the largest open-source AI models currently available.
This scale suggests the model is primarily intended for enterprise and research-grade applications, rather than consumer-level use, due to significant hardware requirements.
While the Hugging Face repository includes core configuration files and model weights, it currently lacks detailed documentation, performance insights, or deployment guidelines.
Licensing and Commercial Use
One key aspect of this release is its MIT licensing, which enables developers, researchers, and businesses to freely use, modify, and integrate the model into proprietary or commercial products.
This move may broaden adoption, especially among enterprise users seeking high-performance, customizable models without restrictive licensing terms.
Performance and Benchmarks
In code generation tasks, R1-0528 has demonstrated competitive performance.
According to LiveCodeBench—a benchmark developed collaboratively by UC Berkeley, MIT, and Cornell—the model ranks just below OpenAI’s o4-mini and o3 models, while outperforming xAI’s Grok-3-mini and Alibaba’s Qwen-3.
Reinforcement Learning Without Supervised Fine-Tuning
A notable feature of DeepSeek’s approach is its training strategy. R1-0528 was trained using reinforcement learning (RL) without an initial supervised fine-tuning (SFT) phase.
This method enables more autonomous learning and enhances the model’s chain-of-thought (CoT) reasoning, allowing for abilities such as self-verification, iterative reflection, and the generation of complex multi-step outputs.
Distilled Versions
To increase accessibility and support academic research, DeepSeek has also released several distilled versions of the R1 model.
These include adaptations based on Llama and Qwen architectures. One version, DeepSeek-R1-Distill-Qwen-32B, has surpassed OpenAI’s o1-mini in multiple benchmarks, achieving new performance highs among dense models in the open-source community.
Open-Source Strategy
The release of R1-0528 aligns with DeepSeek’s broader strategy to offer transparent and accessible AI tools.
The company’s rise has also drawn regulatory attention, particularly from U.S. agencies concerned about the geopolitical implications of advanced AI development outside Western institutions.
While the current update does not introduce radically new features, its availability on Hugging Face reinforces DeepSeek’s long-term commitment to open-source AI. It also signals a steady escalation in global competition over AI capabilities and influence.
