In rapid-fire releases, DeepSeek launched their V3 model in December 2024, followed by R1 in January 2025. Both models match OpenAI's best performers in capability, but that's not what made headlines. What really turned heads was how they did it:
- They claim to have built these models on what amounts to a startup budget (in AI terms)
- They have open-sourced nearly everything, letting anyone peek under the hood
- They managed this despite trade restrictions meant to prevent exactly this kind of advancement by a Chinese lab
The Technical Breakthroughs
What makes DeepSeek special isn't just what it can do, but how it does it. The team at DeepSeek made some clever innovations that change the game:
Pure Reinforcement Learning
DeepSeek matched OpenAI's reasoning capabilities using pure reinforcement learning, skipping the expensive labeled data that others rely on. It's a significant achievement that could reshape how we train AI models.
Data Compression
They change how their AI models store information in two ways. First, they use 8-bit numbers instead of 32-bit ones, cutting memory use to a quarter without losing meaningful accuracy. Second, they found a way to compress their key-value cache by 93% while keeping performance intact.
Multi-Token Prediction
While other models process text one token at a time, DeepSeek processes multiple tokens simultaneously, doubling speed without sacrificing coherence.
Mixture of Experts (MoE)
Their architecture uses specialized subnetworks with a smart routing system, making advanced AI accessible on regular hardware. It's a practical approach that could democratize AI capabilities.
What This Means for Law Firms
For legal organizations looking at DeepSeek, there are four main ways to use it, each with its own security implications:
Direct Use
Using DeepSeek's direct application or API services is a non-starter for law firms and legal departments. Their terms of service clearly state they'll use your data for training - that's an immediate red flag for client confidentiality.
Cloud Provider Hosting
Getting DeepSeek through Azure or AWS offers better security, but you'll need to carefully review the cloud provider's terms and security measures.
Self-Hosted Deployment
Running your own DeepSeek instance gives you maximum control over security and data. But it requires significant technical expertise and resources - it's really only viable for firms committed to building solutions.
Vendor Integration
Some vendors are already incorporating DeepSeek into their products. The key question here is how they're doing it - are they hosting it themselves or using DeepSeek's APIs? The difference matters for security.
Two Persistent Concerns
Two other issues need addressing:
1. Model Bias: DeepSeek's training likely reflects Eastern perspectives and values. This means using it to reason over legal matters may give different outcomes to what you expect from other models. Of course, Western models have their own biases too - they're just different ones.
2. Security Risks: There's speculation about potential backdoors in the model. While it's a serious consideration, it's important to note this assumes both state involvement and the technical feasibility of embedding such capabilities.
Making Smart Choices
As consumers of these language models, your approach to DeepSeek should depend on your specific use cases:
- If you're considering direct implementation, run thorough validation tests. Ask the question, does the model actually give useful outputs in your use case. The cost savings and efficiency gains could be substantial if the output meets your standards.
- If you're evaluating vendors using DeepSeek, in addition to assessing the capability of their software to solve the needs in your use case, you should dig into their implementation details and security measures.
Looking Ahead
DeepSeek represents both opportunity and challenge for legal tech. The efficiency gains and cost savings are tempting, but they need to be weighed against security and privacy requirements. As with any new technology in our field, the key is finding the right balance between innovation and prudent risk management.
The pace of AI development isn't slowing down, and DeepSeek won't be the last surprise in this space. Whatever decision you make today should leave room for adaptation tomorrow.
About the Author
Horace Wu is the CEO of Syntheia. Syntheia gives transactional lawyers the tools they need to move faster and work smarter. The platform helps lawyers quickly review documents, spot key issues, and compare terms across deals — so they always know what’s market and what’s different. Syntheia leverages a range of cutting-edge technologies to streamline deals, so lawyers can focus on what really matters: winning the best terms for clients.