GLM-4.7

GLM-4.7 achieves an overall score of 72.9 percent and carries the speed profile Batch Tool Expert in the Leaderboard. That describes the character of this model quite well: not a nervous sprinter for real-time dialogue, but a broadly capable cloud generalist with a clear lean toward tool-, code-, and workload-oriented tasks that can afford to take their time. For a model positioned as a generalist Frontier system, that is respectable, but not flawless; for a system also conceived as an instruct and coder model, the primary issue is uneven reliability — the factor that keeps good technology from becoming a confident platform. Sovereign Risk: HIGH — GLM-4.7 comes from Zhipu AI in China; processing and vendor jurisdiction are, according to available cards, located in China under PIPL/CSL/DSL, with correspondingly elevated third-country and access risk.

Header Notes: Stability and Reliability

Architecture and Classification

The pre-assigned category fits surprisingly well. GLM-4.7 is classified as a Generalist, but shows clear talent in areas where structured execution matters: code, tool affinity, formatted work outputs. The Instruct label explains why many responses remain visibly disciplined, even though a great deal is happening internally. The Coder tag is likewise not mislabeled. In security-adjacent code tasks, there is real substance, even if it does not extend to every last critical detail. And Thinking-Optional matters here, because the specific test run used mode n/a. There was therefore no explicitly activatable thinking switch in the benchmark — just the standard mode of the cloud endpoint.

Editorially, the calibration across the model class is decisive. GLM-4.7 is a Frontier model in the sense of this benchmark, with 355 billion total parameters, but in a MoE architecture with only 32 billion active parameters per token. That active capacity is the fair benchmark. One should not draw conclusions from the total parameter count alone. MoE is finely balanced division of labor rather than raw force. When such a model appears strong, it is often through specialization and efficiency. When it shows gaps, it does so despite an impressive total size.

There is also the deployment mode to consider: GLM-4.7 ran here as a cloud Open Weights model via OpenRouter. This matters for readers, because the measured speed is not simply an abstract model value — it is always also a benchmark of the specific cloud infrastructure path. The speed profile Batch Tool Expert accordingly stands for more batch-oriented use: longer, structured jobs, tool-related workflows, less suited to rapid back-and-forth chat.

Performance Character: More Pack Horse Than Fencing Champion

The speed profile Batch Tool Expert is not a decorative badge but a useful shorthand for the deployment profile. GLM-4.7 does not work like a quick assistant that jumps in immediately, but like a colleague who first sorts through the documents and then responds. For documentation work, security analysis, transformation tasks, and tool-adjacent workflows, that can work well. For anything that lives in a tight interaction loop, the heavy scatter is a genuine problem.

Token economy also matters here. The model formally stays within all budgets, but frugal is something else entirely. Particularly notable is the output volume in the CLI area, Cultural Intelligence, UX Writing, and Content Transformation. This is not a quality bonus. For a cloud model, it simply means: more text, more cost, more wait time. Anyone using GLM-4.7 productively pays not only for answers, but often also for the tendency to turn a screwdriver request into a full toolbox delivery.

API Cost Profile

GLM-4.7 produces significantly more tokens than the fleet median across several modules. In the CLI Benchmark, the average is 3,877 tokens; the fleet median is 312. That corresponds to a factor of 12.43. In Cultural Intelligence, it is 2,796 tokens against a median of 290, a factor of 9.64. UX Writing also registers with 3,410 versus 1,644, a factor of 2.07, and Content Transformation with 3,215 versus 1,837, a factor of 1.75.

For a cloud endpoint, this is not a trivial matter. It simply means: identical task, significantly more billable text. Because the benchmark separates the quality score from this, one can see very cleanly where performance and efficiency diverge. GLM-4.7 is not wasteful in the sense of pointless verbosity, but it is visibly not a model for minimalists.

Code Quality and Security: Substantial, but No Comprehensive Coverage

The code and security side is the most interesting aspect of GLM-4.7. This is where the model demonstrates why the combination of Generalist, Instruct, and Coder was a reasonable assignment. In a demanding PHP security audit, it identifies 16 of 19 expected vulnerabilities, delivers a clean Markdown table, keeps the brief explanations concise, and provides mostly directly actionable fixes. SQL injection, plaintext passwords, mail header injection, IDOR, path traversal, weak token generation: it lands. The security instinct is there.

But Frontier in this benchmark also means: a good hit rate is not quite enough. The gaps are not decorative — they are substantive. Missing, among other things, are a critical SQL injection in the delete path, CSRF protection as a standalone issue, the risk of hardcoded root credentials, and the absence of an expiry time for reset tokens. That is precisely where GLM-4.7 tips from very good auditor to good auditor with blind spots. In security contexts, that is a distinction that cannot be argued away.

On the positive side, the fixes largely remain practical. The model does not merely say “insecure” but suggests prepared statements, password_hash(), or whitelisting strategies. On the negative side, it lacks synthesis. The golden standard builds an attack chain from individual issues. GLM-4.7 stays more at the line level than at the level of compromise paths. It recognizes the attacker’s toolkit, but not always the full intrusion path.

Stability, however, casts a shadow over this strength. In the code quality module, there were dropouts and very high scatter. In security reviews in particular, that is unwelcome. A model that delivers good hits but not reliably every time is like a penetration tester who is brilliant twice and misses the appointment on the third engagement.

Reasoning and Logic: Correct, but Not with Final Elegance

In logical reasoning, GLM-4.7 displays a pleasant characteristic: it is rarely wrong, but often brief. On the classic guard puzzle, the core logic is sound. The strategy of asking what the other guard would say is correctly derived, the formulation is clear, and the conclusion is right. The model does not fail at the thinking — it fails at the elaboration.

This fits precisely with the Thinking-Optional category in standard mode. Because no explicit thinking mode could be activated in the benchmark, no expansive proofs are to be expected. Nevertheless, it remains visible that a great deal happens internally while only part of it surfaces externally. The Judge protocols fault not so much reasoning errors as missing breadth: alternative formulations, explanation of the double-inversion principle, more pedagogical depth. GLM-4.7 solves the task, but it does not lecture. Sometimes that is a virtue. In a reasoning benchmark, it costs points.

The real problem here, again, lies in practical stability. Dropouts occur in the reasoning module. For users, this means: the model can deliver good logic, but not with the reliability one would expect from a Frontier cloud service.

Content Transformation: Surprisingly Strong, with Minor Scratches

In Content Transformation, GLM-4.7 delivers one of the more convincing performances in the test. A complex German video script with an analysis block, timing, production notes, hook, pattern interrupt, Easter egg, and clear spoken-language style is implemented almost completely. The structure holds, the stage directions land, the dramaturgy works. Here GLM-4.7 does not work like a dry text generator but like a capable producer with a sense of pacing.

Particularly pleasing is the clean language handling. The visible output stays in German, the format requirements are respected, and the script is directly usable. Minor issues such as the unidiomatic “onderschept worden” are exactly what they should be: cosmetic flaws, not structural damage.

For a generalist, that is strong. For an instruct-adjacent Frontier candidate, it is almost more important than some benchmark cosmetics, because it shows that GLM-4.7 does not immediately lose track of multiple simultaneous requirements. Language, format, length, production readiness: it delivers.

Cultural Intelligence and UX Affinity: Solid, but Not Quite at Home

In culturally and tonally sensitive tasks, GLM-4.7 performs better than one might expect from a model that presents as strongly technically oriented. In rewriting a toxic, gender-biased HR text, it cleanly removes problematic terms, keeps the output in German, and respects the instruction to provide no explanations. The result is professional, usable, and linguistically correct.

What is missing is sensitivity in the final stretch. The Judge describes this very aptly: slightly less idiomatic, slightly more formal, slightly less inviting than the reference. It is the difference between “correctly phrased” and “genuinely approachable.” The model rewrites the text, but it does not quite breathe the HR conventions that generate trust in the German-speaking context. Given a vendor whose stated strengths lie explicitly in Chinese-English bilingualism, this nuance is not surprising. It is not a fundamental quality defect, but it marks the boundary between good adaptation and cultural fluency.

CLI, Tool Affinity, and Hallucination Risk: Useful Until the Moment It Counts

The most difficult characteristic of GLM-4.7 surfaces where tool results and their exact evaluation meet. The model carries the speed profile Batch Tool Expert, and its general tool affinity is visible in the score. But the constraint log contains a finding that cannot be minimized: in a tool-use task, a hallucination was detected. The model generated content that did not originate from the retrieved tool result but was fabricated. The score was capped as a result.

For content-critical tasks, this is a warning signal. Anyone working with tool output wants reliable execution, not creative supplementation. This is precisely where GLM-4.7 shows a dangerous reflex: it can construct a plausible answer from an incomplete factual basis. In marketing, that might pass as charm; in research, compliance, or incident analysis, it is poison.

There is also a second systemic finding from the CLI area: in one task, internal reasoning tokens crowded out the output budget. There are no visible reasoning tokens in the product, but the effect is real. The model consumed so much internal compute that too little remained for the actual output. For the user, this looks like a model that runs out of breath halfway up the stairs. Methodologically, this is not a content error. Practically, it is still a product deficiency.

Documentation Quality: Usable, but Not the Comfort Zone

Documentation scores fall below what is known from the stronger documentation models in this class. GLM-4.7 writes comprehensibly and can structure content, but it does not project natural authority there. This fits the overall picture: the model is more work-oriented than style-defining. Those seeking solid structure, good readability, and technical soundness will get something usable. Those who treat documentation as precision writing will find more elegant candidates in this class.

Data Protection and Data Sovereignty

For companies in Germany and the EU, the data protection situation with GLM-4.7 is the second major warning signal after stability. The vendor is Beijing Zhipu Huazhang Technology Co., Ltd., headquartered in Beijing, China. Applicable law according to the vendor card is China (PIPL/CSL/DSL), and the stated data location is China. A verified GDPR DPA is not documented; data retention is factually unclear at -1 days.

The calculated Sovereign Risk is HIGH. The reasoning is concrete: as a Chinese company, Zhipu AI is subject to Chinese security and data law; at the same time, the card explicitly references the BSI warning of 04.02.2025 regarding Chinese AI cloud services as the relevant risk benchmark. For European companies, this does not automatically constitute a usage prohibition, but it does mean a significantly elevated third-country risk. Without clearly documented transfer mechanisms and without a substantive DPA, use involving personal or confidential business data is difficult to justify under data protection law.

There is also a second point, distinct from the deployment question: the weights provenance risk is likewise HIGH. Here, the origin of the model and the hosting risk converge rather than diverge. Anyone using GLM-4.7 via OpenRouter or comparable cloud access should not treat this as a formality. Data sovereignty here is not a marketing term — it is an open flank.

Conclusion

GLM-4.7 is an interesting, capable, and in parts impressively useful Frontier model. As a cloud Open Weights model via OpenRouter, it combines broad generalist capabilities with clear strengths in code, security analysis, and structured transformation. The MoE architecture with 32 billion active parameters goes a long way toward explaining why the model often feels substantial without looking like an absolute top performer in every discipline.

The catch, unfortunately, is not one you can overlook. Practical stability is weak, tail latency is critical, and token economy is costly across several modules. And the hallucination in a tool-use task is not a cosmetic flaw — it is a breach of trust in precisely the wrong area. For batch-heavy technical workflows, code reviews, security scans, and structured text transformations, GLM-4.7 can be genuinely useful, as long as a human is reviewing the output and retries are planned for. For time-critical agent chains, fact-strict tool pipelines, and privacy-sensitive enterprise processes, it is not a relaxed recommendation in its current form. GLM-4.7 is not a pretender. But it is also not yet a model you hand the keys to the engine room without supervision.

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.