DeepSeek V4 Pro 0813
Runs MCP tools
The general-availability release of DeepSeek’s largest V4 model, from August 2026. DeepSeek says its biggest change is agent behaviour in production. You can run it on your MCP tools side by side with the April release.
What DeepSeek V4 Pro 0813 is
DeepSeek V4 Pro is a 1.6T-parameter mixture-of-experts model that activates 49B parameters per token, with a 1M-token window and text-only input. The 0813 checkpoint is its general-availability release. DeepSeek says it greatly improves agent capability, especially in production, adds native support for the OpenAI Responses API adapted for Codex, and offers three reasoning-effort levels: low, high and max. The weights are open under the MIT licence. The April release, 0423, is still available here for comparison.
What DeepSeek says
- Humanity’s Last Exam with tools: 60.0.
- Terminal-Bench 2.1: 87.9.
- DeepSWE: 62.7.
- Greatly enhanced agent capabilities, with the largest gains in production environments.
These are the vendor’s own claims, not measurements of ours. Run the model against your own server to find out whether they hold for your tools.
How it reaches MCP
- DeepSeek’s API as deepseek-v4-pro, OpenAI-compatible, which now always serves the 0813 release
- Codex, through DeepSeek’s native Responses API support. Codex can act as an MCP client.
- Any MCP client over OpenRouter as deepseek/deepseek-v4-pro-0813, which is what runs here
What to watch for
- The names disagree. On DeepSeek’s API, deepseek-v4-pro now means 0813. On OpenRouter, deepseek/deepseek-v4-pro is still the April 0423 release, and the GA one is deepseek/deepseek-v4-pro-0813. Check which you are calling before you compare results.
- DeepSeek’s own numbers put V4.1 Flash ahead of V4 Pro on DeepSWE (74.2 against 62.7) and Terminal-Bench 2.1. Test Flash on your tools before assuming you need Pro.
- Input is text only. Image content returned by an MCP tool, such as a screenshot, will not reach the model. V4.1 Flash reads images if you need that.
Give 0423 and 0813 the same multi-step prompt against the complex-schema mock, for example creating a profile and then calling configure_workflow for it. Agent behaviour is what DeepSeek says changed, and a chain of dependent calls is where it shows. All the mock servers →
Variants available here
| Model | Model ID |
|---|
| DeepSeek V4 Pro 0813 | deepseek/deepseek-v4-pro-0813 |
| DeepSeek V4 Pro 0423 | deepseek/deepseek-v4-pro |
Frequently asked questions
What is the difference between DeepSeek V4 Pro 0423 and 0813?
0813 is the general-availability release. DeepSeek says it greatly improves agent behaviour, especially in production, adds native OpenAI Responses API support, and brings low, high and max reasoning-effort levels. Both are available here, so you can compare them on your own server.
Which DeepSeek V4 Pro does the API return?
On DeepSeek’s own API, the name deepseek-v4-pro serves the latest release, now 0813. On OpenRouter the plain deepseek/deepseek-v4-pro id is still the April release, and the GA release has its own id, deepseek/deepseek-v4-pro-0813.
DeepSeek V4 Pro or V4.1 Flash for MCP?
DeepSeek reports V4.1 Flash ahead on agentic coding (74.2 against 62.7 on DeepSWE v1.1), and it also reads images. V4 Pro is the larger model. Run the same prompt on both against your server. The comparison takes a minute and is the only evidence that matters for your tools.
Does DeepSeek V4 Pro work with MCP?
Yes, through any MCP client that translates tools into function calls. OpenRouter lists tools, tool_choice and structured_outputs for it. Codex can also drive it through the Responses API, and Codex speaks MCP.