DeepSeek-V4-Pro
DeepSeek's open-weight flagship under the MIT license, for teams that need to self-host or want a low-cost hosted API for agent and coding work.
Install
-
Weights
deepseek-ai/DeepSeek-V4-Pro-0813Hugging Face repo ID. Ships without a Jinja chat template; use the encoding scripts in the repo.
-
DeepSeek API model ID
deepseek-v4-pro
The August 2026 release of DeepSeek-V4-Pro, published as open weights under the MIT license. You can run it on your own hardware, fine-tune it, or call DeepSeek’s hosted API.
When to choose it
- Data can’t leave your environment, or you need a model you control and can pin forever.
- You want a permissive license with no usage restrictions for commercial work.
- Cost is the deciding factor for high-volume agent or coding jobs and you use the hosted API.
Specs (from DeepSeek, 2026-09-25)
| Weights | deepseek-ai/DeepSeek-V4-Pro-0813 (MIT) |
| API ID | deepseek-v4-pro |
| Context (API) | 1M tokens |
| Max output (API) | 384K tokens |
| API price | $0.66–$1.32 per million input tokens (cache miss), $1.98–$3.96 output; the lower rate applies off-peak |
| Reasoning effort | low, high, max |
Before you choose
Self-hosting this model needs a multi-GPU server cluster; the weights run to well over a trillion parameters. It’s not something you can run on a laptop. The hosted API is DeepSeek’s own service, so read its data terms before sending anything confidential. API peak hours (higher price) are 01:00–04:00 and 06:00–10:00 UTC on weekdays.
- Publisher
- DeepSeek
- Works with
- Any agent
- Categories
- Agent building, Backend & APIs, Model selection
- License
- MIT (model weights)
- Cost
- Free tier, paid plans