DeepSeek Unveils DeepSeek R1-0528 : What’s New and Performance
Chinese AI startup DeepSeek today released an incremental yet impactful update to its flagship R1 reasoning model, designated DeepSeek R1-0528, on the Hugging Face platform. Published under the permissive MIT license on May 28, 2025, the update builds upon the original R1 release from January 2025, which first demonstrated that open-source language models could rival proprietary offerings in both performance and cost-efficiency . Despite being described internally as a “minor trial upgrade,” R1-0528 incorporates substantive enhancements in reasoning sophistication, code synthesis accuracy, and model dependability .
Update Highlights of DeepSeek R1
Extended Contextual “Thinking”
Leveraging Mixture-of-Experts (MoE) layers and optimized attention mechanisms, R1-0528 can sustain deeper chains of reasoning over prolonged inputs, making it well-suited for workflows that demand extended context retention ([Analytics Vidhya][4]).
Improved Reliability & Consistency
Developers report more stable output across repeated prompts, with reduced variance in answer quality and fewer timeout errors when handling long, context-rich queries.
Enhanced Reasoning
Underlying architecture remains at 671 billion parameters, leveraging the same attention mechanisms as the original deepseek R1 but with refined weights and longer “thinking” capability .
Internal evaluations show measurable gains on logic-intensive benchmarks, with more coherent multi-step inference across complex queries .
Superior Code Generation
LiveCodeBench benchmarks report clear gains in AI-assisted coding: R1-0528 now ranks as the top Chinese-origin model for code tasks, trailing only OpenAI’s o4-mini-high, o3-high, and o4-mini-medium configurations .
Outperforms Alibaba’s Qwen 3 and Anthropic’s Claude 3.7 on coding accuracy and efficiency.
Technical Performance
| Metric/Benchmark | R1-0528 | R1 (Jan 2025) | Leading Proprietary |
|---|---|---|---|
| Composite LLM Score | Median 69.45 | ~67.8 | OpenAI o3/o4-mini (~70–72) |
| Code Generation Rank | #1 among Chinese models; overall #4 worldwide | #2 among Chinese models | Behind OpenAI o4-mini & o3; ahead of Grok 3 mini |
| Context Window | 163 840 tokens (≈2 × original) | 81 920 tokens | Comparable to OpenAI o3-high |
| Inference Efficiency | 37 b active params / 671 b total | Same | Similar to other large open-source |
Benchmark Platforms:
LiveCodeBench: Positions R1-0528 just below OpenAI’s o4-mini and o3 on coding tasks, surpassing xAI’s Grok 3 mini and Alibaba’s Qwen 3.
Composite LLM Aggregates (MMLU, HumanEval, GSM8K, BBH, TruthfulQA): Yields a median performance of 69.45, placing it within striking distance of Claude and Gemini offerings (Analytics Vidhya).
Memory & Context:
Expanded context window supports codebases, long‐form documents, and multi-turn dialogues with minimal loss of coherence.
Open-Source Licensing & Accessibility
Released without a public model card on Hugging Face, the R1-0528 weights and configuration files are nevertheless fully accessible under the MIT license, permitting unrestricted commercial use and community-driven modifications. DeepSeek’s choice of license continues its strategy of “open weights” dissemination, ensuring that research institutions and startups alike can integrate the model without encumbrance.
Market Impact & Competitive Landscape
DeepSeek’s R1 launch in January 2025 disrupted prevailing assumptions about the cost and scale required for cutting-edge AI, prompting U.S. rivals to adjust pricing and introduce lighter models (e.g., OpenAI’s o3 Mini) . With R1-0528, DeepSeek reasserts its position as the leading open-source contender, driving competition on both performance and price dimensions against OpenAI, Google Gemini, Alibaba, and Anthropic.
Future Outlook
Since its January launch, DeepSeek R1 has already prompted strategic responses from major AI players—most notably influencing OpenAI’s pricing adjustments and Google’s roadmap for Gemini models. The R1-0528 update is expected to intensify competitive pressure, particularly among firms seeking cost-effective yet high-performance alternatives to proprietary LLMs . Meanwhile, U.S. regulators have raised national security concerns over DeepSeek’s rapid advances, underscoring the geopolitical stakes of open-source AI proliferation.
Getting Started
CometAPI provides a unified REST interface that aggregates hundreds of AI models—under a consistent endpoint, with built-in API-key management, usage quotas, and billing dashboards. Instead of juggling multiple vendor URLs and credentials.
Developers can access the latest DeepSeek R1 API (model name: deepseek-r1-0528) through CometAPI.To begin, explore the model’s capabilities in the Playground and consult the API guide for detailed instructions. Before accessing, please make sure you have logged in to CometAPI and obtained the API key, price in CometAPI:
- Input Tokens: $0.44 / M tokens
- Output Tokens: $1.752 / M tokens
