DeepSeek-V4-Pro Exposed as "One API, Three Models"? Community Testing Reveals Discrepancies Possibly Due to Agent Environment

August 15th

Recently, the AI community has been discussing the suspicion that there are multiple versions of DeepSeek-V4-Pro. Some users have discovered that when calling the deepseek-v4-pro endpoint, switching IP addresses, or recreating sessions, the model exhibits three different "inference styles":

One type frequently starts with "Let me," similar to the previous V4 Pro Preview; another often says "The user wants me," resembling V4 Flash; and a third type uses "we" extensively, dubbed by some users as the more powerful "God Edition V4 Pro."

As once a session enters a certain mode, its subsequent behavior usually remains stable, the community speculated that DeepSeek may have hidden multiple models behind the API and distributed them through a routing mechanism. However, further analysis of the DeepSeek Harness source code has led to another explanation: the differences may not stem from different model weights but from the Agent runtime environment.

The community discovered that on August 10th, the DeepSeek Harness official repository received a crucial commit:

"fix(preset): align minimal agent with RL composition."

This update aimed to align the Minimal Agent with the Agent environment used during Reinforcement Learning (RL) training.

According to the official documentation, the Minimal preset includes a minimal system prompt, persistent Bash environment, specified editing tools, compaction policy used in RL training, and removes additional identity prompts, web prompts, and tool descriptions.

This suggests that DSH Minimal may not be a "watered-down" version of the Standard edition but rather an Agent environment encountered during simulated model training.

Community tests also support this view. The same DeepSeek V4 Pro performs differently in various Harness environments:

DSH Standard: 91 points;

DSH PTC: 92 points;

DSH Minimal: 99/96 points.

Subsequently, the tester developed the "Anchored Standard" plugin: the initial request simulated a Minimal environment, only opening the shell and read tools. After completing the initial tool invocation, the full Standard toolset was restored, resulting in consistently scoring 98/99.

The tester believes that the key to unleashing the V4 Pro Agent's capabilities may not lie in the final number of tools it possesses, but rather in what the model sees initially: System Prompt + Tool Schema + Agent Scaffold.

Therefore, the so-called "three DeepSeek models" may actually be a two-layered stacking of factors:

On the one hand, there are differences in API service environments, deployment configurations, or gray instance variations; on the other hand, it is whether the model enters an Agent environment close to the RL training distribution.

However, this assertion has not yet been confirmed by DeepSeek officials. According to the official API documentation, deepseek-v4-pro corresponds to the DeepSeek-V4-Pro-0813 production version, and the mechanism for automatic routing of multiple models has not been disclosed.

Currently, it appears that the different performances of DeepSeek-V4-Pro are more likely the result of a combination of model weights, inference environment, and Agent framework, rather than simply the existence of three hidden models.