Deepseek-R1-0528: BEST Opensource Reasoning Model! Powerful, Fast, & Cheap! Fully Tested + Free API

Deepseek R1 0528 Model Release Summary

Key Concepts:

• Deepseek R1 0528: An upgraded open-source reasoning model by Deepseek AI.
• Sparse Mixture of Experts (MoE): Architecture leveraging 671B total parameters with 37B active for inference.
• MMLU (Massive Multitask Language Understanding): A benchmark for evaluating large language models.
• Chain of Thought Reasoning: A method where the model breaks down a problem into a series of logical steps.
• Counterfactual Reasoning: The ability to reason about what might have happened if circumstances were different.
• API Access: Accessing the model through Deepseek's API or Open Router.
• Token Cost: Cost per million input and output tokens.

1. Model Overview and Release Details:

• Deepseek AI has released Deepseek R1 0528, an upgraded open-source reasoning model.
• The model utilizes a sparse Mixture of Experts (MoE) architecture with 671 billion total parameters, but only 37 billion parameters are active during inference, enhancing efficiency.
• The model was released on Hugging Face under the MIT license.
• Initial reports suggest improved long-term reasoning and suitability for real-world applications compared to the original Deepseek model.

2. Benchmark Performance:

• The model's performance on the MMLU benchmark is comparable to models like Grok-1.5 Mini and Grok-1.5 High.
• It outperforms several established models, including Grok-1.5 Mini and the 2024-0131 update of Grok-1.5 Mini, in MMLU.
• It demonstrates competitive general reasoning ability, especially considering it's an open-source model competing with proprietary models.
• The model also performs well in hard question accuracy.

3. Access and Cost:

• Access is available through Deepseek's API or a free API from Open Router (with a paid version for unlimited access).
• The context window is approximately 136K tokens.
• Pricing is $1.95 per 1 million input tokens and $5 per 1 million output tokens.

4. Prompt Testing and Results:

• Math Reasoning:
  • Prompt: A train travels at 60 mph for 45 minutes, then 30 mph for 30 minutes. What is the total distance traveled and the average speed?
  • Purpose: To test multi-step arithmetic reasoning, unit conversion, and attention to detail.
  • Result: The model correctly calculated the average speed (48 mph) and the distance for the first segment (45 miles), demonstrating chain-of-thought reasoning and numerical consistency.
  • Generation time: 106.88 seconds.
• Coding and Creativity:
  • Prompt: Draw a beautiful sunset skybox suitable for an early 2000s Sega video game.
  • Purpose: To test coding and creative generation capabilities.
  • Result: The model generated a "pretty decent job" in generating the skyline.
• SAS Landing Page Generation:
  • Prompt: Create a SAS landing page with as many features as possible.
  • Purpose: To assess front-end design and modern landing page creation abilities.
  • Result: The model generated a modern-looking landing page with animations and highlighted features, including pricing plans. The result was described as "really beautiful".
• Common Sense and Counterfactual Reasoning:
  • Prompt: If it's raining and John didn't take his umbrella, what is most likely the outcome? What could John have done differently?
  • Purpose: To evaluate common sense reasoning and hypothetical thinking.
  • Result: The model correctly identified the likely outcome (John getting wet) and provided counterfactual reasoning, suggesting John should have checked the weather, carried an umbrella, or found shelter. It also identified the effect of the rain, the absence of protection, immediacy, as well as probability.

5. Key Arguments and Perspectives:

• The Deepseek R1 0528 model represents a significant performance jump compared to its predecessor, despite being a minor upgrade.
• The model's open-source nature and competitive performance against proprietary models make it a valuable resource.
• The model's reasoning capabilities and performance on various tasks suggest it is well-suited for real-world applications.

6. Conclusion:

The Deepseek R1 0528 model is a notable advancement in open-source reasoning models. Its improved performance, efficient architecture, and reasonable pricing make it a compelling option for various applications. The successful prompt testing across different domains highlights its versatility and potential for real-world use. The release is likely a step towards the development of the Deepseek R2 model.