What the evidence says
These facts come from the reviewed Epoch AI catalog artifact and its linked primary record. They describe the model. They do not rank it.
- Developer [1]
- DeepSeek
- Released [1]
- April 24, 2026
- Domains [1]
- Language
- Tasks [1]
- Language modeling/generation · Question answering
- Access [1]
- Open weights (unrestricted)
- Weights [1]
- Open
Where it may fit
These are evidence-bounded screening suggestions, not performance claims or purchase recommendations.
Self-hosted language evaluation
Consider it when open weights and unrestricted access are requirements for a language-model evaluation.
Evidence fields: domains, access, openWeights [1]
Question-answering prototypes
Consider it for question-answering research that benefits from inspectable, deployable weights.
Evidence fields: tasks, access, openWeights [1]
Limitations
The record does not provide hardware requirements, serving cost, context window, latency, safety evaluations, or a directly matched Arena score. Deployment suitability still requires independent testing.
Current standing
According to the LMArena leaderboard dataset published July 12, 2026, DeepSeek-V4-Pro holds a rating of 1457.1 in the style-controlled text category, with a reported 95% interval of 1452.6 to 1461.6 across 41,800 battles. This is one source's point-in-time observation, not a composite score or a claim of overall superiority.
See the full leaderboard · LMArena source ↗
Head-to-head comparisons
Source-backed pair pages that put DeepSeek-V4-Pro next to another documented model. None declares a winner beyond what a named source's numbers say.