Pryvan
Model

DeepSeek

DeepSeek's open-weight models are known for strong reasoning and code performance with efficient mixture-of-experts architectures, a capable, cost-effective option to host inside the EU.

Efficient reasoning and code

DeepSeek

  • Strong reasoning and code
  • Efficient mixture-of-experts design
  • Open weights
  • Cost-effective at scale
Model family

Available variants

DeepSeek-R1 Distill 70B

Reasoning-optimized (Max)

DeepSeek-R1 Distill 32B

Reasoning that fits Core

DeepSeek-Coder

Specialized for code

Benchmarks

How it performs

Indicative published results for DeepSeek-R1 Distill 70B. Scores vary by version, harness and prompt, confirm for the version you deploy.

General knowledge MMLU86
Graduate reasoning GPQA57
Coding HumanEval87
Math MATH-50088
Model card

What you're deploying, in plain English.

Trained on

Trained on a large corpus with heavy code and reasoning content. Details vary by version.

Strengths

  • Reasoning and problem-solving
  • Code generation and review
  • Compute efficiency at scale

Limitations & biases

  • Behaviour and tone differ from Western models
  • Standard hallucination risk
  • Validate outputs for your jurisdiction and policies

Recommended use cases

  • Developer assistance
  • Analytical and reasoning tasks
  • Technical documentation

Prohibited / high-risk

  • Unreviewed high-risk decisions
  • Use beyond your defined lawful basis

At a glance

Vendor
DeepSeek
Languages
Multilingual, strong in English and code
License
Open weights
Context
Long context
Sizes
Available in multiple variants
Hosted in the EU. Runs on vanafter compute or your own infrastructure, your data never leaves European jurisdiction.
Deploy on Pryvan

Best for: Reasoning- and code-heavy work that needs to stay cost-efficient.

One arrow. One direction. Forward.

Bring AI into your business, without giving up your data.

Join the waitlist. We're onboarding GDPR-sensitive SMEs across Europe.