Hook
On August 15, the AI community erupted with a conspiracy theory: DeepSeek-V4-Pro is not one model but three. Users reported that calling the deepseek-v4-pro API from different IPs or fresh sessions yielded three distinct inference styles—one starting with “Let me”, another with “The user wants me”, and a third heavily using “we” (dubbed the “God Version”). The narrative spread fast: DeepSeek is hiding multiple models behind a routing algorithm, A/B testing on unsuspecting developers. But as someone who has spent the last decade auditing codebases from ICOs to DeFi to AI-crypto oracles, I’ve learned one rule: Code doesn’t lie. The truth is far more subtle—and far more revealing about the state of AI agent engineering.
Context
DeepSeek-V4-Pro is the latest flagship from the Chinese AI lab, positioned as a direct competitor to OpenAI’s GPT-4o and Anthropic’s Claude 3.5. The model is accessible via a single API endpoint, deepseek-v4-pro, which the official documentation claims corresponds to the “DeepSeek-V4-Pro-0813” release. There is no mention of multi-model routing or A/B testing. Yet the community’s observations were systematic: users on X (formerly Twitter) and Discord documented that once a session entered a particular mode, the inference style remained stable. They hypothesized that DeepSeek was deploying a “mixture of agents”—maybe a fast, cheap version, a balanced version, and a premium version—and routing users based on random allocation or load balancing.
This is not a new phenomenon in AI. OpenAI has been caught routing queries to different model versions without disclosure. But the crypto-AI intersection has taught me to be skeptical of such claims. In 2026, when I investigated the convergence of AI and blockchain oracles, I saw how easily the community conflates environmental artifacts with deliberate model selection. The DeepSeek case is a textbook example.
Core
The real story is not multiple models—it’s the Agent environment. On August 10, the official DeepSeek Harness (DSH) repository received a critical commit: fix(preset): align minimal agent with RL composition. The commit message is dense but the implication is clear: the “Minimal Agent” preset must be identical to the environment used during reinforcement learning (RL) training. Let me break down what that means.
DeepSeek Harness is a benchmarking framework used to evaluate the agentic capabilities of DeepSeek models. It has several presets: Standard, PTC, and Minimal. The Minimal preset strips away all identity prompts, web prompts, and tool descriptions. It provides only a minimal system prompt, a persistent Bash shell, a set of specified editing tools, and a compaction policy—exactly the environment the model saw during its RL training phase. The official documentation states that "the Minimal preset simulates the real Agent environment encountered during model training."
Based on my audit experience—from the 2017 ICO reality checks to the 2024 Bitcoin ETF legal deep dive—I know that when a system’s behavior changes across identical inputs, the first place to look is the environment, not the weights. A neural network is a function of both its parameters and the context in which it runs. The system prompt, the tool schema, the agent scaffold—these are not just decoration. They are active constraints that shape the model’s output distribution.
Community tests confirmed this. The same DeepSeek V4 Pro model was evaluated across different DSH environments: - DSH Standard: 91 points - DSH PTC: 92 points - DSH Minimal: 99/96 points
That’s a 8-point swing purely from changing the environment. The model’s intrinsic weights did not change. The difference is the prompt and tool setup.
Then came the killer experiment: the “Anchored Standard” plugin. Testers designed a custom environment where the first request mimicked the Minimal environment—only shell and read tools, no web access. After the first tool call, the full Standard toolset was restored. The result? Consecutive scores of 98/99 points. The model performed as if it were in the Minimal environment, even though it later had access to more tools. The key insight: the first impression is what matters. The initial system prompt and tool schema anchor the model’s behavior. Once the model is in “agentic mode” from the RL training distribution, it can handle additional tools without losing performance. But if the model starts with a full Standard environment—with identity prompts, web prompts, and extraneous tool descriptions—it gets confused and performs worse.
This is why users saw three inference styles. The API infrastructure likely has multiple deployment instances, some running with a full Standard agent scaffold, some with a Minimal-like scaffold, and some with a mixed configuration. The “Let me” style corresponds to the Standard environment where the model was trained to output a cautious preamble. The “The user wants me” style is the Minimal environment where the model is in pure agent mode. The “we” style might be a third variant—perhaps a different system prompt or a different compaction policy. These are not different models. They are the same model loaded into different agent environments.
Contrarian
The community’s insistence on “three hidden models” is a classic case of confirmation bias meets Occam’s razor. It’s easier to believe in a conspiracy of multiple models than to understand the nuance of RL training distribution alignment. But there is a more dangerous blind spot: the assumption that model performance is a direct function of the weights alone. This is the same mistake that led DeFi traders in 2020 to believe that high APY meant protocol sustainability. In reality, the source of the yield (or in this case, the performance) depends on the environment—the context, the scaffold, the first prompt.
From my 2026 AI-crypto convergence report, I warned that centralized AI providers could hide performance variability behind opaque APIs. If DeepSeek can change the inference style by changing the environment, then any claim about “benchmark scores” is meaningless without knowing the exact environment used. This has implications for decentralized AI inference networks. If a model is deployed on-chain or via a DAO, the smart contract must specify the exact agent environment—not just the model hash—to guarantee reproducible behavior. Otherwise, users could get different results depending on the validator’s configuration.
Furthermore, the Anchored Standard experiment reveals a potential attack vector: if the first request determines the behavior, an adversary could craft a malicious initial prompt that forces the model into a vulnerable mode. In a blockchain oracle context, where models are used to verify real-world data, this could be catastrophic. A model that starts in Minimal mode is highly capable but also more susceptible to adversarial instructions because it lacks guardrails. The compaction policy and minimal system prompt remove the “identity” that prevents the model from executing dangerous commands. The RL training environment may have been designed for maximum efficiency, not safety.
Takeaway
The DeepSeek-V4-Pro saga is not a story of hidden models. It is a story of environmental determinism. The same model, same weights, different behavior. The crypto-AI space must learn this lesson: auditable inference requires more than model transparency; it requires environment transparency. The next time you see a community screaming “multiple models,” look at the agent scaffold first. Code doesn’t lie—but it does require the right context to tell the truth. The question DeepSeek needs to answer is not “Are there three models?” but “Why are there three different agent environments in production for a single API endpoint?” Until they explain that, the real risk is not confusion—it’s the illusion of consistency.