DeepSeek-R1: Why This Open-Source AI Model Matters

Author(s): Paul Ferguson, Ph.D.

Originally published on Towards AI.

New AI models emerge almost weekly, making it hard to distinguish between the significant improvements and the minor updates. DeepSeek-R1, however, represents a clear exception.

While its performance matches or slightly exceeds leading proprietary models (like OpenAI’s o1) in many tasks, there are three reasons why this model is important:

1. Cost Efficiency: Trained for only 5–10% of the cost of comparable models
2. Open Accessibility: Fully open-source under an MIT licence
3. Technical Innovation: Novel methods like self-taught reasoning with task-focussed processing

What makes this model so interesting isn’t just its performance but how it’s achieved: its open-source framework and over 90% cost reduction put pressure on closed systems to innovate while enabling businesses to deploy advanced AI affordably.

In summary: its combination of efficiency, transparency, and adaptably set a new benchmark for the industry.

Competitive Performance Without the Premium Price

Independent benchmarks show DeepSeek-R1 performs comparably to closed models in a number of domains: this challenges the assumption that open-source AI lags behind proprietary systems:

While it’s marginally behind in general knowledge (e.g., MMLU: 90.8% vs. 91.8%), it has clear advantages in technical tasks: making it particularly suited for software engineering, financial modelling, and scientific research.

Open-Source Design

Closed models require costly API subscriptions, whereas DeepSeek-R1’s MIT licence allows for:

• Full customisation: Modify the model for niche applications (e.g., healthcare,, legal contract analysis, etc).
• Local deployment: Smaller variants (1.5B–70B parameters) run on consumer-grade GPUs (avoiding cloud fees). In a previous article, I discussed the growing importance of Small Language Models, and some of these models fit neatly into that category).
• Transparency: Independent audits of model weights to address bias or safety concerns.

Novel Methods

DeepSeek’s cost and efficiency advantages come within three main areas:

Reinforcement Learning First

• Self-taught reasoning: Learns through trial and error problem solving rather than expensive human feedback
• Discovery phase: Explores new strategies (e.g., it will attempt to verify its own answers)
• Alignment phase: Refines outputs for coherence and accuracy

Predicting Two Steps Ahead

• Training: Forecasts the next two tokens at once
• Inference: Produces answers faster through parallel token prediction

Sparse, Task-Specialised Processing

• Only 5.5% of parameters (37B/671B) are activated per query

Cost Savings

DeepSeek’s pricing changes what businesses can achieve with limited budgets:

• Free to use via its web app.
• Although for business use cases, these are typically carried out through API calls
• Provides API access at a relatively low cost ($0.14 for 1 million input tokens, compared to $7.5 for OpenAI’s o1 model)
• To companies with significant usage of LLMs, these differences can add up to thousands of dollars over the course of a month

Implications

1. Democratisation: Smaller companies can more easily compete with larger businesses.
2. Pressure on Closed Models: Companies like OpenAI are under pressure to reduce their prices or increase the transparency of their models.
3. Ethical Trade-Offs: Though open weights help in bias mitigation, unregulated customisation risks misuse.

Conclusion

DeepSeek-R1 proves that AI progress does not have to rely on closed systems or unsustainable compute budgets.

For organisations, this means faster experimentation, lower barriers to entry, and control over AI tools: a combination likely to accelerate innovation across different industries.

While not flawless, its open-source model and technical ingenuity set a new standard for what’s possible in efficient, accessible AI.

If you’d like to find out more about me, please check out www.paulferguson.me, or connect with me on LinkedIn.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI