BlockBeats reported on Aug. 15 that the AI community has been debating whether DeepSeek-V4-Pro is showing up in more than one version behind the same API.
Users said that when calling the deepseek-v4-pro endpoint, switching IP addresses or starting a new session could lead the model to display three different 「reasoning styles」.
One style often begins with 「Let me」 and was seen as close to the earlier V4 Pro Preview. Another frequently uses 「The user wants me」 and was described as similar to V4 Flash. A third makes heavy use of 「we」 and has been labeled by some users as a stronger 「god-tier」 V4 Pro.
Because a session usually keeps behaving the same way once it enters one of those modes, some community members initially speculated that DeepSeek might be hiding multiple models behind the API and distributing traffic through a routing mechanism.
A code update pointed to another explanation
After a closer look at the DeepSeek Harness source code, a different explanation began to take shape. The variation may not come from different model weights. It may come from the Agent runtime environment instead.
Community members pointed to a key commit posted to the official DeepSeek Harness repository on Aug. 10: 「fix(preset): align minimal agent with RL composition」. In the community's reading, that update was meant to make the Minimal Agent consistent with the Agent environment used during reinforcement learning, or RL, training.
According to the official documentation, the Minimal preset includes a minimal system prompt, a persistent Bash environment, designated editing tools, and the compaction policy used in RL training. It also removes extra identity prompts, web prompts, and tool instructions.
Under that interpretation, DSH Minimal may not be a cut-down version of Standard. It may be an attempt to mirror the actual Agent environment the model encountered during training.
What the tests showed
Community testing was also cited as support for that view. The same DeepSeek V4 Pro reportedly produced different scores under different Harness environments:
- DSH Standard: 91
- DSH PTC: 92
- DSH Minimal: 99/96
Testers then built an 「Anchored Standard」 plugin. On the first request, it simulates the Minimal environment and exposes only shell and read tools. After the first tool call is completed, it restores the full Standard tool set.
According to the tester, that setup produced consecutive scores of 98/99.
The key variable may be the first Agent scaffold the model sees
The tester argued that the main factor behind V4 Pro Agent performance may not be how many tools the model has in the end. The more important factor may be what the model sees first: the System Prompt, the Tool Schema, and the Agent Scaffold.
From that perspective, the appearance of 「three DeepSeek models」 may actually reflect two layers of variation at once. One layer would be differences in API service environment, deployment setup, or canary instances. The other would be whether the model is placed into an Agent environment that is closer to the RL training distribution.
No official confirmation of multi-model routing
That theory has not been confirmed by DeepSeek. The official API documentation shows that deepseek-v4-pro maps to the production release DeepSeek-V4-Pro-0813, and it does not disclose any automatic multi-model routing mechanism.
Based on what is publicly available for now, the differing behavior of DeepSeek-V4-Pro appears more likely to reflect the combined effect of model weights, inference environment, and the Agent framework, rather than the simple existence of three hidden models.