GLM-4.7

With an overall score of 73.97%, GLM-4.7 enters the field as a peculiar hybrid: a generalist server model with an instruct focus, a clear affinity for code, and an optional thinking architecture — tested here in fixed default mode without a thinking toggle (n/a). The speed profile badge “Batch DevOps Expert” reveals a great deal about its character: not a sprinter built for hectic dialogues, but a model for longer, tool-adjacent workflows where throughput matters more than instant response. For a Cloud Open-Weights model via OpenRouter, this is a respectable showing — but not a flawless one. Sovereign Risk: HIGH: Zhipu AI / Z.AI is a Chinese company and subject to China’s National Security Law (NSL), which may allow state access to data. In February 2025, Germany’s BSI explicitly warned against the use of Chinese AI cloud services (BSI reference: Warning DeepSeek, 04.02.2025); this risk assessment applies analogously to all Chinese cloud AI providers that process user data on Chinese servers. With purely local inference, the Cloud Act-equivalent risk does not apply.

Header Grades: Stability and Reliability

Architecture and Expectations

The pre-assigned category General, Instruct, Coder, Thinking-Optional fits surprisingly well. GLM-4.7 aims to be many things at once: an all-rounder for broad tasks, direct instruction-following, solid technical grounding, and deeper reasoning on demand. The specific test mode is decisive here: n/a, meaning default behavior without a separately toggled thinking run in the benchmark. This matters because some of its responses show traces of internal depth without presenting visible chain-of-thought notes to the reader. What was measured, therefore, is not the theoretically maximum deliberative version, but the variant a typical API user actually receives.

The model classification also sets the right benchmark. GLM-4.7 is a Generalist, not a pure specialist model. It belongs to the Size Class Server and uses a MoE architecture with 355.0 billion total parameters, but only 32.0 billion active parameters per token. That is precisely the standard against which it must be measured. The total size sounds heavyweight; the active capacity is more like a controlled mid-range entry in the server tier. Anyone expecting raw Frontier-class power is misreading the architecture. Anyone hoping for specialization, efficiency, and solid breadth, however, is on the right track.

Performance Profile: Batchable, Not Nimble

The badge “Batch DevOps Expert” is more than decoration. It describes GLM-4.7 quite precisely: a model that is more at home in batch-oriented, technical workflows than in fast-paced, conversational ping-pong situations. Generation quality feels moderate to somewhat sluggish, and for a Cloud Open-Weights model via OpenRouter, this assessment is always also a benchmark of the provider and its infrastructure — not merely of the weight set itself. These speeds should therefore be read as characteristics of the deployed cloud endpoint.

The downside shows up brutally in the long tail. On average, things may still feel manageable, but the outliers run so long that productive chains of tool calls, agent logic, and user interaction grind to a halt. A model is allowed to be thorough. It is allowed to be slow. But it must not routinely behave as though time has no cost.

Code Quality and Security: Solid Foundation, Missing Audit Bite

The Code Quality score of 67.08% describes GLM-4.7 fairly accurately. It spots the obvious problems, names SQL injection, XSS, session fixation, path traversal, weak tokens, and several implicit security gaps. This is not blind fumbling. The model understands the landscape. It simply lacks, at the critical moments, the paranoid instinct that good security reviews require.

The qualitative log is telling. In one security task, GLM-4.7 found 15 out of 19 relevant vulnerabilities. The omissions that carry the most weight are CSRF protection, missing expiry on reset tokens, and a sensitive header ordering issue. These are not footnotes. These are the classic spots where “messy” quickly becomes “exploitable.” Similarly, the model tends to stay too generic on implicit vulnerabilities. It names type juggling but does not explain the concrete attack vector with the sharpness one would expect in a professional security review. The same applies to fixes: usable, but not hardened. === is mentioned; hash_equals() is not. basename() appears; realpath() validation does not. Cookie flags are touched on; SameSite is left behind.

Precisely because GLM-4.7 runs as a Coder, this is not a side issue. A coder-adjacent server model need not perform magic, but a robust security radar is a reasonable expectation. Here it delivers the profile of a competent developer rather than an uncomfortable auditor. Useful for initial analyses. Not sufficient for sign-offs.

CLI and Tool-Adjacent Practice: Strong on Technical Tone, Weak on Tool Fidelity

In the CLI benchmark at 86.67%, GLM-4.7 shows one of its clearer strengths. The model appears comfortable in the DevOps register, follows technical task specifications mostly directly, and delivers actionable responses. This aligns with the instruct side of its architecture and with the badge. Anyone seeking shell-adjacent assistance gets not a prose-polishing model but a worker with a toolbox.

However, this section also contains the report’s most serious trust violation. In a tool-use task, a hallucination was identified: the model generated content that did not originate from the retrieved tool output but was fabricated. The score was capped accordingly via the hallucination cap. For content-critical tasks — research, fact-bound summarization, or agent chains with external data sources — this is a disqualifying signal. Because that is precisely where the hardest rule applies: when a model treats tool outputs not as ground truth but as loose inspiration, assistance quickly becomes fiction with a protocol timestamp.

This also drags down the ToolUse score to 42.5%. GLM-4.7 is thus considerably stronger at direct technical responses than at strictly source-bound tool execution. That is a distinction that must not be blurred in practice.

Reasoning and Logic: Correct, But Not Elegant

At 73.55% in logical reasoning, GLM-4.7 lands in solid territory. The qualitative picture is almost textbook: the model arrives at the correct solution, explains the relevant cases coherently, and remains linguistically clean. In the classic guard puzzle, for instance, it produces the correct question, correctly analyzes both guard scenarios, and arrives at the appropriate course of action.

What it lacks is didactic precision. The judges rightly note that the model suggests the double inversion rather than genuinely working it out. It solves the task but does not make the elegance of the solution visible. That is exactly what separates “correct” from “good” in this module. For practical answers, this is usually sufficient. For a model with optional thinking depth, it is nonetheless a minor flaw. It apparently thinks enough to reach the destination, but not always in a way that lets the reader benefit from the journey.

The framing matters here: GLM-4.7 is architecturally a Thinking-Optional model, but this benchmark ran in default mode. Shorter, more direct answers are therefore not a failure of the setup but the chosen operating state. The potential for depth is visible. Its best pedagogical form is not always delivered.

UX Writing: Usable, But Not Yet Senior-Level

In the UX Writing & Microcopy module, GLM-4.7 reaches 76.29%, revealing a familiar model character: it can optimize, tighten, and clarify. But when psychological depth, prioritization, and completeness converge, competence no longer becomes excellence.

The judge log describes the performance aptly as junior-level. The fundamentals are there: jargon is reduced, the mobile perspective is considered, tone is improved. But on psychological principles the response remains underdeveloped, important parts of a multi-step task are left incomplete, and a metrics table is missing entirely. This is not catastrophic. It is, however, the kind of gap that frustrates product teams in day-to-day work because it forces a second pass. In this area, GLM-4.7 delivers not a misstep but half-finished usability. You keep working with it. You do not simply sign it off.

Documentation Quality and Content Transformation: Well-Structured, Sometimes Too Schematic

Documentation Quality at 74.69% and Content Transformation at 74.91% reveal one of the model’s more appealing sides. GLM-4.7 can organize material, cast it into reliable forms, and structure longer outputs cleanly. In the transformation module in particular, the judge appreciated that the model completed a full three-part task in clean German, including analysis, production script, and a built-in Easter egg. The result was production-ready and in some details even stronger than the reference style.

What stands out, however, is the tendency toward functional structure over stylistic elegance. In the video script, for example, GLM-4.7 chose a tabular format rather than a more cinematic screenplay layout. This is not wrong — it is often more practical from an editorial standpoint. But it reveals something about the model’s character: it prefers to think in columns, lists, and work surfaces. Less in dramaturgy.

Cultural Intelligence: Linguistically Confident, Culturally Decent, But Not Maximally Nuanced

At 73.84%, GLM-4.7 delivers a solid performance in Cultural Intelligence. Particularly encouraging is its linguistic discipline. In the available logs, the model maintains clean German throughout, follows clear specifications, and avoids unnecessary explanatory loops. In a task involving a more inclusive rewrite of a job posting, it reliably removes toxic language and phrases things idiomatically.

The weakness lies in the final calibration. The model defuses a great deal, but not everything. Phrasings such as “Durchsetzungsvermögen” and a competitive tone remain somewhat too present in places. The judge describes this accurately as a good but not excellent result: professional, readable, functional — just not quite at the level of modern, genuinely warm and inclusive HR language. It is the linguistic equivalent of a correctly fitted suit that is missing its final tailor’s crease.

API Cost Profile

GLM-4.7 is a cloud model with metered usage, and for that reason its verbosity is not an academic side note but a pricing question. In the Cultural Intelligence area, the model produces an average of 2,166 tokens against a fleet median of 257 — 8.43 times the average across all tested models. In the CLI area, it generates 1,488 tokens versus 283, or 5.26 times the median. In UX Writing, it produces 3,768 tokens against a fleet median of 1,676, or 2.25 times. Content Transformation also comes in at 1.9 times the median, with 3,477 versus 1,832 tokens.

This needs to be stated plainly: this additional length is not inherently a quality gain. In several areas, GLM-4.7 handles tasks adequately while producing significantly more text than necessary. For API usage, this simply means higher costs for comparable utility. Anyone integrating the model into batch processes should look not only at response quality but at the end-of-month bill. At $0.40 per million input tokens and $1.75 per million output tokens, GLM-4.7 is not expensive in absolute terms. But its verbosity erodes the cost advantage faster than one might like.

Data Privacy and Data Sovereignty

For European organizations, GLM-4.7 is not a relaxed candidate from a data protection standpoint. According to the Vendor Card, Zhipu AI is headquartered in Beijing, China. Chinese law (PIPL/CSL/DSL) applies, the data location is China, and no verified status exists for a GDPR DPA. The data retention period is listed as -1 days, making it factually non-transparent.

The calculated Sovereign Risk is HIGH. The rationale is concrete: Zhipu AI is a Chinese company; additionally, the card references the BSI warning of 04.02.2025 regarding Chinese AI cloud services as an analogous risk reference. For users in Germany and the EU, this means: anyone sending personal, confidential, or regulatorily sensitive data to this endpoint is operating in a third-country setup with elevated access and compliance risk. Without clearly documented transfer mechanisms and a solid contractual basis, this is not a peripheral concern for many enterprise scenarios — it is a disqualifying criterion.

Conclusion

GLM-4.7 is a capable, sometimes impressively structured server model with genuine technical substance. Its profile as a Generalist with an instruct and coder lean is credible: CLI tasks, structured transformation, and many direct technical assignments suit it well. The MoE architecture with 32 billion active parameters also explains why the model does not present itself as a pure top-tier Frontier system, yet often delivers more than a merely adequate mid-range stack. It has character, and in this field character is rarely a disadvantage.

The problems, however, are too real to paper over with architectural romanticism. The instability with 7 timeouts, the critical tail latency, the significant verbosity cost burden, and above all the hallucinated tool output meaningfully limit confidence. In security tasks, GLM-4.7 thinks usably but not rigorously enough. In UX and cultural tasks, it writes decently but not finely enough. Anyone looking for a versatile batch assistant for technical support work can operate with GLM-4.7. Anyone needing a model for strictly fact-bound tool chains, sensitive enterprise data, or unattended production pipelines should know very precisely why they are accepting that risk. Across all tests, hallucination is the decisive flaw precisely because it does not occur constantly — but surfaces at exactly the moment where it is least forgivable.

This evaluation was generated automatically based on the benchmark data. Model used: GPT-5.4 by OpenAI. The raw data and full methodology are documented in the GitHub project.