DeepSeek V4 Pro

With an overall score of 75.05%, DeepSeek V4 Pro makes its intentions unmistakably clear: a Frontier reasoning model with MoE architecture, broad strategic reach, and considerably more seriousness than charm. The assigned speed profile badge Interactive Tool Expert fits well: this Cloud Open-Weights model is not optimized for showroom speed but for deliberate, usable responses in interactive tooling and analysis workflows. As a reasoning-optimized Frontier model with 1.6 trillion total parameters but only 49 billion active parameters per token, it plays the card of specialized capacity rather than raw full-activation. Sovereign Risk: HIGH — DeepSeek operates under Chinese jurisdiction; per a BSI warning dated 02/04/2025, the cloud service is not recommended for official or sensitive data.

Header Grades: Stability and Reliability

The architectural classification is not a footnote here — it is the key to understanding the model. DeepSeek V4 Pro simultaneously carries the labels Thinking, Thinking-Optional, Agentic-Orchestrator, and Long-Context. At first glance this sounds like a collision of product ideas, but in the benchmark it works with surprising coherence: the model visibly thinks less like a short instruct sprinter and more like a planner that structures tasks rather than firing them off reflexively. It is important to note that the optional deep thinking mode was deliberately not activated during benchmarking. What was measured is the default behavior — exactly what a typical API user gets without any special configuration.

Performance and Character in API Operation

DeepSeek V4 Pro was tested here as a Cloud Open-Weights model via OpenRouter. This is central to contextualizing the speed figures: the measured 40.54 tokens per second is not an abstract value for the model in a vacuum, but the result of the provider’s deployed cloud infrastructure including the network path. For readers, this means: the speed describes the quality of the provided endpoint in everyday use, not some theoretical model speed.

The badge Interactive Tool Expert says more than a pretty label. It describes a model intended for interactive work with close tool proximity: analyses, structured outputs, step-by-step assistance, technical support. That is precisely where DeepSeek V4 Pro tends to feel at home. It is not nervous, not frantic, not drilled for maximum brevity. But the P95 response time of 87.96 seconds also means: in a noticeable share of requests, you wait long enough that the system’s deliberation stops feeling like a virtue and starts feeling like an interruption. For an Agentic-Orchestrator, this is not entirely surprising. For productive, time-critical loops, it remains a cost in patience.

Reasoning and Logic: Thought Through Correctly, Not Always Delivered Correctly

For a model with a reasoning focus, logical performance is the first test. DeepSeek V4 Pro reaches 72.01% here. That is not a misstep, but it is not a triumph either. Qualitatively, a recurring pattern emerges: the model often arrives at the correct solution but does not always explain it with the depth its category promises. In the guardian riddle, for instance, it delivers the correct key question and therefore the right answer. The Judge rightly notes, however, that the actual explanation remains conspicuously brief. A model that wants to be read as a Thinking system cannot get away with a correct final sentence and nothing more.

This point matters specifically for categorization. A Thinking model is allowed to be longer. It is allowed to be more thorough. In logic tasks, that is precisely what it is supposed to do. When a correct result arrives but the derivation shrinks to instruct-level, it feels like a sports car stuck in third gear.

Metacognition Compliance (Reasoning): The model refuses to use the explicitly requested <thought> tags in 3/5 metacog tests, citing a consistent policy statement. The reasoning content itself is partially correct — the score deduction results from format non-compliance, not from errors in thinking. CrucibleMark deliberately evaluates native zero-shot instruction compliance as a real-world characteristic. This deduction is methodologically intentional.

There is also a documented language error in this same module. In one reasoning task, the model ignored the explicit German-language requirement and responded in English. The system scored this as a Language Mismatch. The penalty here is not an aesthetic objection but a clean application of the rules: when a target language is explicitly prescribed, a response in the wrong language is simply unusable in production.

In one reasoning task, the model additionally violated the explicit German-language requirement. The system did not apply a discretionary style deduction — it registered a formal rule violation. The content quality of the response is therefore secondary. Anyone using reasoning in multilingual workflows receives a clear warning signal here.

Code Quality and Security: Analytically Strong, But Not Without Gaps

At 67.88% in Code Quality, DeepSeek V4 Pro falls below the level one would reasonably expect from a reasoning-heavy Frontier model with security ambitions. This is not a total failure. It is more the case of a capable auditor who misses four critical points on the whiteboard and thereby squanders credibility at precisely the most important moment.

The qualitative record shows this precisely. In a security analysis, the model identifies 15 out of 19 relevant vulnerabilities. That is decent, but not sufficient in a real security audit. What is particularly critical is not some cosmetic omission, but the failure to flag findings such as hardcoded secrets, missing CSRF protection, or non-expiring reset tokens. Equally problematic is the tendency to underrate the severity of individual findings. When IDOR or Path Traversal is treated as high rather than critical, what is missing is not knowledge but a sense of escalation. That is precisely where security understanding parts ways with security vocabulary.

It must also be said in DeepSeek V4 Pro’s favor that its fix suggestions are usually precise and actionable. The table structure holds. The wording is clean. The analysis does not feel hallucinated but deliberately prioritized. The model does not fail here through chaos but through incompleteness. That is the more dangerous weakness, because it looks professional.

For use as a security assistant, this means: suitable for initial analyses, useful for structured audit lists, but not reliable enough for the final sign-off. Anyone using it to review source code should plan for a second reviewer — not out of distrust of the model’s intelligence, but of its tendency to treat the last 20 percent as optional.

CLI, Tooling, and Agentic Workflows

The CLI benchmark at 86.67% is one of the model’s stronger areas. This fits the Agentic-Orchestrator attribution. DeepSeek V4 Pro thinks in task structures, not just individual output sentences. That is precisely what helps with shell, tool, and workflow tasks, where planning is often more important than a single brilliant one-liner.

The category deserves to be read with appropriate fairness. Agentic-Orchestrator models are conductors rather than solo violinists. When a model plans, decomposes, safeguards, and plausibly sequences tool steps, that is often more valuable in real agent setups than perfect exact-matching on every individual line. DeepSeek V4 Pro delivers the right kind of competence in this domain: controlled technical orientation with high tool proximity. It does not aim to impress — it aims to organize the workflow. That is a good quality. It is just less glamorous than spectacular bullseyes.

Documentation Quality: Solid, Dense, Usable

At 78.75%, DeepSeek V4 Pro delivers one of its most well-rounded performances in documentation quality. This is not surprising. Models with long context windows and a reasoning-oriented design can organize large bodies of material more effectively, provided they maintain the discipline not to drown in it. DeepSeek V4 Pro mostly maintains that discipline.

Its strength here lies less in literary elegance than in traceable structure. The model explains, connects, and organizes reliably. It does not write with the flair of a top editor, but with the dependability of a technician who understands that a good text must above all be usable. Especially for internal documentation, technical guidelines, or process descriptions that require explanation, this is a meaningful advantage.

Content Transformation and UX Writing: Creative Enough, But Not Always Obedient

In Content Transformation, the model reaches 76.55%. The qualitative example of a reworked video script illustrates clearly where DeepSeek V4 Pro can shine: it delivers complete structures, timestamps, visual cues, production notes, and a convincingly rendered spoken tonality. This is not accidental text output but a workable script. It is apparent that the model can hold complex requirements in mind simultaneously.

But there is a formal thorn here as well. In one Content Transformation task, the model exceeded the explicit word limit of 250 words, producing 358 words — 143% of the limit. The system applied an automatic deduction of 20%, or 16.32 points, to the achieved score. The content quality of the response is irrelevant to this. The penalty applies regardless. Anyone working in marketing, editorial, or workflow automation with hard length limits should take this seriously. The model can write. But it can treat word count constraints with noticeably more latitude than the prompt permits.

In UX Writing at 78.63%, a related trait emerges. DeepSeek V4 Pro often produces usable, differentiated, and reasonably nuanced microcopy. However, it is not a model of the concise point. When turned loose on a small copy task, it tends to respond with the enthusiasm of an intern eager to demonstrate that they understood every idea. That can be pleasant in creative workshops. In everyday API use, it primarily generates higher token costs.

Cultural Intelligence: Solid, But Not Infallible

The score of 70.64% in Cultural Intelligence is solid but not outstanding. The model can defuse tones, clean up problematic phrasing, and produce inclusive language reliably. The qualitative record on revising a toxic job posting demonstrates exactly this. DeepSeek V4 Pro meaningfully replaces aggressive or male-coded language, keeps the output in German, and largely maintains the required HR tone.

The interesting detail lies in the nuance: the solution is functionally convincing but not always the most elegant linguistically. When the model opts for inclusive forms with visible markers, that is formally acceptable but not necessarily the most stylistically refined choice. It does the right thing, but sometimes with somewhat blunter tools. That is better than cultural blindness. It is just not yet mastery.

API Cost Profile

DeepSeek V4 Pro is priced affordably, but it is not economical in expression. Precisely because it is used as a Cloud Open-Weights model via OpenRouter, this distinction is practically relevant: more output tokens mean directly higher API costs at identical quality.

The overhead is particularly noticeable across several modules. In the CLI domain, the model produces an average of 987 tokens against a fleet median of 251 — 3.93 times the average across all tested models. In Cultural Intelligence, it generates 693 tokens versus a median of 219, or 3.16 times as many. In UX Writing, DeepSeek V4 Pro produces 4,236 tokens against a fleet median of 1,493 — 2.84 times the median. Documentation Quality at 4,353 vs. 2,877 tokens and Code Quality at 3,989 vs. 2,526 tokens further confirm that the model becomes more verbose than necessary almost reflexively.

This is not a quality judgment. DeepSeek V4 Pro delivers good results in several areas. It simply talks considerably longer than average to do so. Anyone billing by API usage will not find a bargain here — just a low-rate plan with a very talkative participant.

Long Context: Ample Space, Sensibly Used

The context window of 1,000,000 tokens is not marketing decoration but a genuine characteristic of the model. This benchmark does not directly probe the full million, of course. Nevertheless, documentation, transformation, and structure-heavy tasks show that DeepSeek V4 Pro does not merely manage longer contexts — it translates them into coherent outputs. In combination with the MoE architecture, this is particularly interesting: the enormous total size of 1.6 trillion parameters sounds spectacular, but the decisive figure is the 49 billion active parameters. Performance should be calibrated against that active capacity. And measured against it, the results are strong. Not magical. But strong.

Data Privacy and Data Sovereignty

The data privacy situation is the model’s most significant strategic liability. The calculated Sovereign Risk is HIGH. The rationale: DeepSeek is a Chinese company subject to Chinese law — specifically PIPL/CSL/DSL and, relevant in the context of weights provenance, the National Security Law. For European users, this represents a genuine third-country transfer risk without an EU adequacy decision.

The provider card lists China plus EU/US cloud partners as the data location. This sounds more flexible but does not resolve the underlying issue. The provider’s jurisdiction remains determinative. For companies in Germany and Europe, a particularly critical point is that no publicly documented GDPR DPA exists. This is not a cosmetic flaw but a concrete compliance obstacle for GDPR-sensitive deployments. The data retention period is listed as -1 days, making it factually unclear. For professional procurement, this is not a minor detail but a red flag.

The distinction between open weights and cloud operation is also relevant. The weights are openly available under the MIT License. This helps with usage rights. It does not, however, mitigate the weights provenance risk, which is explicitly rated HIGH here. In other words: open does not automatically mean sovereign. With DeepSeek V4 Pro, especially not.

Conclusion

DeepSeek V4 Pro is a Frontier model to be taken seriously, with a clearly recognizable profile. It thinks in structured ways, performs strongly in CLI and documentation tasks, often delivers pleasingly complete results in Content Transformation, and puts its MoE design to sensible rather than merely impressive use. At the same time, it falls short of its own ambitions when reasoning is correct but not articulated with sufficient depth, when security audits miss important gaps, or when formal requirements around language and word count are not cleanly observed.

For practical deployment, the verdict is therefore: highly interesting for analytical knowledge work, technical assistance, documentation-heavy workflows, and agentic tool orchestration. Less suitable for highly regulated enterprise environments, hard format automation without post-review, and any processing of sensitive data in the DeepSeek cloud configuration. Across all tests, no notable hallucinations — the model would rather over-explain than fabricate. That is a respectable trait. It is just a shame that data sovereignty and instruction discipline are precisely where the finish noticeably scratches.

This evaluation was generated automatically based on the benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and full methodology are documented in the GitHub project.