The AI community has been buzzing with a conspiracy theory: DeepSeek is hiding multiple versions of its V4-Pro model behind a single API endpoint. Users reported that by rotating IP addresses or recreating sessions, they could elicit three distinct 'inference styles' — one that starts with 'Let me', another that declares 'The user wants me', and a third, more verbose variant dubbed the 'God Version'. The logical conclusion was that DeepSeek had deployed a routing mechanism that dynamically selects different model weights. But as usual, the crowd jumped to the most exciting conclusion without auditing the plumbing. I’ve spent the last week tracing the actual code paths, and the reality is far more mundane — and far more instructive.
DeepSeek-V4-Pro is supposed to be a single unified model, released as the 0813 checkpoint. The official API documentation states that deepseek-v4-pro corresponds to that exact version, with no mention of multi-model routing. Yet the community’s empirical evidence is undeniable: different sessions produce different behavior patterns. The phenomenon is not random; once a session enters a particular mode, it stays there. This suggests a stateful component is influencing the model’s output, not a random lottery of weights.
The key to understanding this lies in the DeepSeek Harness (DSH) source code. On August 10, the official repository updated a commit with the message: fix(preset): align minimal agent with RL composition. This commit ensures that the Minimal preset — a configuration option for the agent environment — matches exactly the environment used during reinforcement learning (RL) training. The Minimal preset strips away identity prompts, web tools, and tool descriptions, leaving only a minimal system prompt, a persistent Bash shell, and a set of editing tools. It is not a 'stripped-down' version of the Standard preset; it is a different simulation of the RL training distribution.
Community tests have confirmed that the same DeepSeek V4 Pro model yields dramatically different scores across Harness environments: DSH Standard scored 91 points, DSH PTC 92 points, and DSH Minimal scored 99/96 points. The difference is not subtle. More tellingly, testers created an 'Anchored Standard' plugin: the first request simulates the Minimal environment (only shell and read tools), and after the first tool call, the full Standard toolset is restored. This approach yielded consecutive scores of 98/99 points. The implication is clear: the model’s performance is gated by the initial system prompt and tool schema it encounters, not by the total number of tools available.
The bottleneck is not the model weights, but the agent scaffold. The model is trained in an RL environment where it learns to interact with a specific set of tools and prompts. When deployed in a different environment — even one with more tools — the model suffers from a distributional shift. The initial prompt and tool schema act as a 'context anchor' that sets the model’s behavior for the entire session. If the anchor matches the RL training distribution, performance is high. If not, the model falls back to a less optimal policy.
Audit the code, not the pitch. The community’s speculation about multiple models is a classic case of mistaking behavioral variation for architectural complexity. The real story is about inference environment alignment, not hidden model versions. DeepSeek’s API likely uses a stateful routing mechanism that assigns sessions to different inference instances — each with a different Harness preset configuration. Some instances run the Minimal preset, others run Standard, and perhaps some run an older PTC preset. The model weights are identical, but the agent environment differs. This explains why the behavior is session-stable but inconsistent across users.
Complexity hides risk. The fact that the model’s performance can swing by 8 points based solely on the system prompt is a red flag for anyone relying on this API for production workloads. It means that the same request can yield different quality of output depending on which instance it hits. This is not a bug; it is a feature of the current deployment architecture. But it introduces a form of non-determinism that is opaque to the end user. The official documentation does not disclose this, which is a failure of transparency.
Let me be clear: I am not accusing DeepSeek of deception. The company is simply running a large-scale inference service with multiple backend configurations, likely to optimize for cost and latency. The Minimal preset is cheaper to run because it reduces tool calls. The Standard preset is more expensive but provides more functionality. The routing algorithm may be based on load balancing or user tier. But the absence of documentation means that users are essentially flying blind, attributing differences in output to model intelligence rather than to environmental quirks.
Trust no one, verify everything. I have reproduced the community’s findings using my own test harness. I created a script that sends the same prompt to the DeepSeek API from three different IP addresses, each with a fresh session. The results matched the reported pattern: one session started with 'Let me', another with 'The user wants me', and the third used a more formal tone. I then modified the requests to include a custom system prompt that mimics the Minimal preset. The variation disappeared. This confirms that the differences are driven by the prompt, not the model.
The contrarian angle here is that the bulls — the community members who claimed there were multiple models — were actually right about the symptom but wrong about the cause. The variation is real, and it does affect user experience. But the solution is not to demand a single unified model; it is to demand consistent environment configuration. DeepSeek could fix this by exposing a preset parameter in the API, allowing users to explicitly choose the Minimal, Standard, or PTC environment. This would give users control over the trade-off between performance and functionality.
Sharding is easy; consensus is hard. In the blockchain world, we often talk about the difficulty of achieving consensus across distributed nodes. The same principle applies here: achieving consistent behavior across millions of inference requests is a technical challenge. The problem is not that DeepSeek is hiding models; it is that they are not hiding the environment mismatch. The community should be pushing for transparency in the deployment configuration, not chasing phantom model versions.
From a regulatory perspective, this issue is relevant to the upcoming MiCA framework for AI systems. The EU’s AI Act will require providers to disclose the intended deployment environment and any known performance variations. If DeepSeek continues to operate with opaque routing, they may face compliance challenges in Europe. The crypto industry has taught us that obscurity is not a sustainable strategy. The same applies to AI.
Based on my audit experience, this is a classic case of a system that works well in controlled conditions but fails to account for deployment variability. The RL training environment is a lab; the production environment is a battlefield. The gap between them is where the bugs live. DeepSeek’s engineers know this — that’s why they wrote the fix(preset) commit. They are trying to align the production environment with the training environment, but they have not yet completed the job. The three 'styles' are gradients of alignment, not separate models.
Looking forward, the key metric for any AI API provider should be 'environmental consistency'. Users should be able to query the same model from different endpoints and get the same behavior. This is not a trivial requirement, but it is a necessary one for trust. The blockchain community has already learned this lesson: consensus requires determinism. AI inference needs the same.
Takeaway: Do not attribute to malice what can be explained by environment misconfiguration. The three DeepSeek-V4-Pro models are a mirage. The real story is about the fragility of agent performance under distribution shift. The fix is not a new model version; it is a better deployment pipeline. And until DeepSeek publishes their environment configuration, the community will be left guessing. That is a failure of both engineering and communication.
Code does not lie, people do. The commit history tells the truth. The only question is whether DeepSeek will act on it.