Hook
On August 15, the AI community erupted over a peculiar anomaly: the same API endpoint, deepseek-v4-pro, was returning three distinct 'inference styles.' Some sessions began with 'Let me,' others with 'The user wants me,' and a third group heavily used 'we.' The narrative quickly split into two camps: one claimed DeepSeek was hiding multiple models behind a single API, routing users based on IP or session; the other, more paranoid, suspected a gradient of capabilities—a 'God Version' reserved for the chosen few. I traced the gas trails back to the root cause, but this time, the gas was not on Ethereum—it was in the API logs. The real story, as I discovered by digging into the DeepSeek Harness source code, is far more subtle and far more instructive for anyone building on decentralized execution environments. The difference is not in the model weights, but in the environment the model sees first. This is a lesson blockchain developers should take to heart: the initial state determines everything, and the code does not lie, but the auditor must dig.
Context
DeepSeek, a Chinese AI research lab, released DeepSeek-V4-Pro in early August 2024. The model is a large language model with advanced reasoning capabilities, and the company offers an API under the name deepseek-v4-pro. According to the official documentation, this corresponds to the DeepSeek-V4-Pro-0813 version, a single model with no disclosed multi-model routing. However, the community's empirical tests revealed systematic differences in behavior. Users reported that once a session entered a particular mode, the style remained stable, suggesting a persistent latent factor. The speculation centered on three possibilities: (1) multiple model versions behind a load balancer, (2) a dynamic routing mechanism that selects a model based on user tier or query complexity, or (3) a bug in the inference pipeline. The community's forensic analysis quickly turned to the DeepSeek Harness (DSH) repository, a testing framework for evaluating model performance in agentic tasks. On August 10, a key commit was pushed: fix(preset): align minimal agent with RL composition. This commit was the smoking gun—but not in the way most assumed. It pointed not to a hidden model, but to the environment's role in shaping the model's output.
In the blockchain world, we are no strangers to such mysteries. I recall auditing a DeFi protocol where the same smart contract function returned different gas costs depending on the transaction order. The root cause was not a different contract, but the state of the EVM at the moment of execution. Similarly, here, the model's behavior is not a function of its weights alone, but of the environment—the system prompt, the tool schema, and the agent scaffold—that precedes the first interaction. This is a critical insight: in both AI and blockchain, the context of execution can be as important as the code itself.
Core
The DeepSeek Harness repository defines several 'presets' for testing the model in agentic tasks. The Standard preset includes a full system prompt, identity instructions, web search tools, and a detailed tool description. The Minimal preset, on the other hand, is described as having a minimal system prompt, a persistent Bash environment, a specified set of editing tools, and a compaction policy that mirrors the environment used during reinforcement learning (RL) training. The key commit ensured that the Minimal preset's agent environment was exactly aligned with the RL training environment. This is not a 'stripped-down' version of Standard; it is a simulation of the exact conditions under which the model learned to use tools. The RL training distribution likely involved a simple shell, limited tools, and a specific prompt format. When the model is deployed in a Standard environment, it encounters a different prompt, additional tools, and identity cues—this distribution shift degrades performance. The Minimal environment, being closer to training, yields higher scores.
Community tests corroborated this. On the same DeepSeek V4 Pro model, scores varied dramatically across Harness environments: DSH Standard scored 91 points, DSH PTC (likely a variant) scored 92, and DSH Minimal scored 99 or 96. The spread is not marginal; it is a 8-point gap, which in agentic benchmarks is massive. The community then developed an 'Anchored Standard' plugin: the first request simulated the Minimal environment, only opening shell and read tools, and after the first tool call, the full Standard toolset was restored. This cross-pollination achieved consecutive scores of 98 and 99 points. The implication is clear: the model's performance depends on the first interaction—the system prompt and the initial tool schema. The model learns to 'settle' into a mode based on that initial context. If the context matches the RL training distribution, the model performs well; if it deviates, the model struggles to adapt.
This is exactly analogous to the concept of 'state' in blockchain. In Ethereum, a smart contract's behavior can change based on the storage state at the time of call. The same function call can produce different results if the state has been mutated by prior transactions. Similarly, the model's 'state' is initialized by the system prompt and tool list. The first tool call then sets the trajectory for the entire session. The community's discovery of three 'styles' is not evidence of three models, but of three different initial states: one that triggers the 'Let me' pattern (likely the Standard preset), one that triggers 'The user wants me' (perhaps a variant with a different prompt), and one that triggers 'we' (possibly the Minimal preset). The code does not lie, but the auditor must dig. The Dig into the DSH commit reveals that the Minimal preset is not a variant; it is the ground truth. The 'God Version' is simply the model operating in its native environment, unfiltered by additional instructions.
Contrarian
The community's immediate assumption—that DeepSeek was hiding multiple models—is a classic blind spot. We see this in blockchain all the time: when a node behaves differently, we assume a different client or a fork, but often the cause is a configuration difference, a gas limit, or a block reorg. In this case, the real blind spot is the assumption that the model is a static entity. Models are not code; they are probabilistic functions sensitive to input distribution. The 'multiple models' narrative is attractive because it is simpler—it requires no understanding of the training environment or the inference pipeline. But it is wrong. The smarter contrarian take is that the API provider is not hiding anything; they are simply failing to align the deployment environment with the training environment. This is a systemic risk, not a deliberate deception.
In my experience auditing Layer 2 systems, I've seen similar patterns. For example, Optimism's sequencer had a bug where the gas estimation for the same transaction differed between the sequencer and the L1 node because of different EVM configurations. The community assumed the sequencer was using a different contract, but the root cause was a discrepancy in the basefee calculation. Shifting the consensus layer, one block at a time, we realized that the environment—the block context—was the variable. The same applies here. The model's 'consensus' on how to answer depends on the initial prompt. The community's tests with the Anchored Standard plugin prove that the environment is the key. The contrarian angle is not just about DeepSeek; it is about the broader lesson for AI agents on blockchain. If we deploy AI agents on-chain, their behavior will be determined by the environment in which they are initialized. A smart contract that calls an AI agent must ensure that the agent's environment matches the training distribution, or the results will be unpredictable.
Another blind spot: the assumption that more tools always improve performance. The Standard preset includes web search, identity, and additional tools, yet it scores lower than Minimal. This is counterintuitive. In blockchain, we often assume that more data or more functionality improves the system, but sometimes it introduces noise. The Minimal environment, with fewer tools, forces the model to rely on its core reasoning. This is reminiscent of the 'less is more' principle in smart contract design: the simplest contract is often the most secure. The DeepSeek evidence suggests that for agentic tasks, the initial tool set should be a subset of the training tool set, and any additional tools should be introduced gradually, if at all.
Takeaway
The DeepSeek-V4-Pro API mystery is not about hidden models; it is about environment consistency. For blockchain developers building AI agents, the lesson is clear: the first interaction is the most important. The system prompt, the tool schema, and the agent scaffold must be carefully aligned with the training distribution. Do not assume that the model will generalize; design the environment to match the training. The code does not lie, but the auditor must dig. In the chaos of a crash, the data remains silent—but here, the data is loud. The next time your AI agent on a blockchain fails to perform as expected, look not at the model weights, but at the environment. The gas trails lead to the root cause, and the root cause is always the initial state. Shifting the consensus layer, one block at a time, we must understand that the environment is the new consensus. The future of on-chain AI will depend on developers who treat the inference environment as a critical component, not an afterthought.
