**[DeepSeek-V4-Pro Allegedly Features "One API, Three Models"? Community Testing Reveals Agent Environment Differences Behind the Scenes]** BlockBeats News, August 15 — Recently, the AI community has been abuzz with discussions about DeepSeek-V4-Pro potentially having multiple versions. Some users have discovered that when calling the deepseek-v4-pro API, changing the IP or recreating a session results in three distinct "inference styles": one frequently starts with "Let me," resembling the earlier V4 Pro Preview; another often includes "The user wants me," similar to V4 Flash; and the third heavily uses "we," which some users have dubbed the more powerful "God-tier V4 Pro." Since a session typically maintains a stable behavior once it enters a specific mode, the community speculated whether DeepSeek might be hiding multiple models behind the API and distributing them via a routing mechanism. However, further analysis of the DeepSeek Harness source code has led to another explanation: the differences may not stem from varying model weights but rather from the Agent runtime environment. The community discovered that on August 10, the official DeepSeek Harness repository updated a critical 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 streamlined system prompt, a persistent Bash environment, specified editing tools, the compaction policy used in RL training, and removes additional identity hints, web prompts, and tool descriptions. This suggests that the DSH Minimal version may not be a "stripped-down" version of the Standard version but rather a simulation of the Agent environment encountered during the model's training phase. Community testing supports this perspective. The same DeepSeek V4 Pro exhibited different performances across various Harness environments: - **DSH Standard:** 91 points - **DSH PTC:** 92 points - **DSH Minimal:** 99/96 points Subsequently, testers developed an "Anchored Standard" plugin: the first request simulated the Minimal environment, only enabling shell and read tools, and after the initial tool invocation, restored the full Standard toolset. The results consistently achieved 98/99 points. Testers believe that the key to unlocking V4 Pro Agent capabilities may not lie in the number of tools ultimately available but rather in what the model first encounters: **System Prompt + Tool Schema + Agent Scaffold.** Thus, the so-called "three DeepSeek models" may actually be the result of two overlapping factors: differences in API service environments, deployment configurations, or gray-scale instances; and whether the model enters an Agent environment resembling the RL training distribution. However, this theory has yet to be confirmed by DeepSeek officially. According to the official API documentation, deepseek-v4-pro corresponds to the DeepSeek-V4-Pro-0813 official version, with no mention of an automatic multi-model routing mechanism. For now, the varying performances of DeepSeek-V4-Pro are more likely the result of the combined effects of model weights, inference environments, and Agent frameworks, rather than the existence of three hidden models. [Original Link]