DeepSeek-R2
Open-weight test-time compute reasoning model with verified formal proofs and mathematics capability.
Source: DeepSeek announcement · Updated 10 May 2026 · provider announcement
Capabilities
Capability ratings
Editorial assessment of what the system can do on each axis, 0-100. Capability is not adoption and not employment impact.
- Code94
Summary
Autonomous theorem proving, complex algorithm optimization, and deep scientific reasoning.
Strongest axis for this model.
Average across all rated axes.
Impact
Occupations affected by DeepSeek-R2 (1)
Occupations with published evidence mapped to this model. A model being able to do something is not evidence that firms use it.
task: Editing and proofreading frontier capability
High overlap between model capability strengths and core occupation tasks.
Provenance
Where this record came from
- Announcement
- https://api-docs.deepseek.com/news/r2
Release records cite the provider announcement. Open-weight releases also link to their public model and repository pages where available. Read /methodology for how capability ratings are produced.