DeepSeek says DeepSeek-V4-Pro-0813 has replaced the company’s V4-Pro preview release and is available across its app, web service and API. The accompanying Hugging Face model card says the weights are released under the MIT License.
DeepSeek has made its V4-Pro model generally available, replacing the previous preview version with the DeepSeek-V4-Pro-0813 release.
DeepSeek said in its API documentation that V4-Pro is now available as a general-availability offering through the company’s app, web service and API. Developers using DeepSeek’s hosted service can access the model through the deepseek-v4-pro identifier, according to the company’s update page.
The release-specific model repository on Hugging Face is named DeepSeek-V4-Pro-0813. DeepSeek’s model card describes that checkpoint as the version replacing the V4-Pro preview release.
The model is also listed by OpenRouter, a third-party platform that provides access to models from multiple providers. That listing indicates that the named release is being offered beyond DeepSeek’s own hosted interfaces, although provider-level availability and operational settings can differ.
DeepSeek’s updates documentation is the primary source for the change in release status. It says V4-Pro has reached general availability and identifies deepseek-v4-pro as the model name for API access.
DeepSeek’s Hugging Face model card provides the most specific documentation for the 0813 checkpoint. It identifies the release as the successor to the preview model, includes DeepSeek’s reported results for agent-oriented evaluations, and states that the weights are available under the MIT License.
Those benchmark figures should be understood as results published by DeepSeek under its documented evaluation setup. They offer a reference point for the company’s claims about agentic capabilities, but they do not establish performance across every application, tool environment or deployment configuration.
OpenRouter’s page is useful as an independent distribution reference for the model name. For details such as model behavior, benchmark methodology and licensing, however, DeepSeek’s own documentation and model card remain the more authoritative sources.
Moving a model from preview to general availability can matter to developers seeking a clearer production target. DeepSeek’s documentation gives hosted users a single model identifier for the general-release version rather than a preview-labelled endpoint.
The MIT licensing statement is also notable for teams evaluating self-hosted or customized deployments. A permissive license can reduce barriers to experimentation and adaptation, but organizations should still review the full license text, relevant laws and their own security and governance requirements before deployment.
DeepSeek’s emphasis on agentic evaluations reflects wider interest in language models that can operate with tools and complete multi-step tasks. Benchmark scores can help inform comparisons, but production decisions also depend on factors including reliability, latency, cost, context handling, safety behavior and performance with the tools used in a specific workflow.
Developers should look for further DeepSeek documentation on pricing, context limits, API behavior and differences between hosted access and downloaded weights. It will also be important to determine whether third-party providers expose the same model configuration and apply their own limits, routing policies or usage terms.
Teams considering a migration from the preview release should independently test the general-availability model with representative prompts, tools and operational constraints before moving production workloads.
DeepSeek has made its V4 Pro model generally available, replacing the previous preview version with the DeepSeek V4 Pro 0813 release.
What happened DeepSeek said in its API documentation that V4 Pro is now available as a general availability offering through the company’s app, web service and API.
Developers using DeepSeek’s hosted service can access the model through the deepseek v4 pro identifier, according to the company’s update page.
Continue reading
