Skip to content
CrucibleMark

GLM-4.7

GLM-4.7 is Zhipu AI’s flagship model with 355 billion total and 32 billion active parameters in a MoE architecture, optimized for agentic coding, reasoning, and bilingual tasks in Chinese and English. The model supports a switchable thinking system and is available as an Open Weights variant for local deployment or via cloud interfaces.

Zhipu AI Version 4.7 Commercial use permitted MoE 355 B (32 B active) 128 K Context 12/2025 $0.38 / $1.74 per 1M

  • Restricted Weights
  • Frontier
  • OR
  • Text
  • Instruction-Tuned
  • Batch

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.

Tool-use profile: 6 assets in detail

Comparison of asset performance (P1/P2/Combined) against the fleet average

Asset performance (radar)

Score Breakdown vs. Fleet Average

Tool-use details

Asset performance, reliability, and runtime profile

Name P1 P2 Combined
EU License Research 80 20 44
HTTP Fetch & Extract 80 35 57.5
Tool Failure Handling (404) 80 60 73
Web Search & Tool Selection 100 60 82
URL Construction & Fetch 80 60 70
Multilingual Search & Synthesis 80 20 50

CrucibleMark evaluates tool use across 6 independent tests. Click a test name for details.

Reliability

  • Tool Call Valid: No
  • Retry: Not required
  • Hallucination: Detected

Reliability measures how consistently a model actually executes tool calls: Tool Call Valid schema and format accepted, Retry Required successful only after retry, Hallucination Flag fabricated tools or parameters detected. All three green means production-ready.

Runtime profile

Call 1
12.9
First Request
MCP
1.44
Protocol Latency
Synthesis
21.16
Response Generation
Total
212.95
Sum of All Phases
Token
6671
Input + Output
Cost
$0.0043
Cost per Run

The runtime profile shows the latency and cost metrics for the model run: Call 1 First Request, MCP Protocol Latency, Synthesis Response Generation, Total sum of all phases, supplemented by Token input and output and Cost cost per run.

Tool-use review

· Instruction-Tuned

Deployment Verdict

Conditional deploy, because GLM-4.7 produces valid tool calls and appears stable in MCP workflows, but synthesis fidelity at Combined 62.75 and detected hallucinations fall short of what is needed to hand it unsupervised trust-critical tool pipelines.

Tool Execution Profile

The operational tool side is the clear strength. At P1 83.33, the model produces valid calls and shows no protocol issues; no retry was necessary. On the Web Search and Tool Selection test, which checks whether the model selects search over fetch without being prompted, it scores P1 100. This suggests genuine tool selection rather than mere schema-following. On the URL Construction test, which measures the derivation of a target URL from the model’s own knowledge, it drops back to P1 80. It can translate familiar patterns into fetch calls usably, but not deterministically enough for fragile endpoints. For MCP orchestration this is workable; for strictly URL-sensitive automations, only with guardrails.

Synthesis Fidelity

How well does it condense tool results? Weakly and inconsistently. P2 42.50 is the real bottleneck. Particularly critical are EU License Research at P2 20 and Multilingual Search & Synthesis at P2 20. The model retrieves information but loses reliability when condensing, prioritizing, and precisely grounding output back to source content. For production systems this means: retrieval works more often than dependable evaluation does.

Does it stay within tool results or fall back on training? In the Honeypot EU License Research test, which checks whether current license restrictions are actually fetched from web sources, no hallucination was flagged — but Content Verification State B1 and P2 20 indicate only weak content binding to the retrieved sources. At the same time, globally hallucination detected: true is a security risk. Once a model outputs fabricated facts as the result of a tool pipeline, trust in the entire infrastructure erodes — even if the tool calls were formally correct.

Error Resilience

On the 404 test, which checks for transparent behavior when tool calls fail, GLM-4.7 remains acceptable. P2 60 is not a strong result, but it does not hallucinate page content despite the error. That is decisive for production. A failed retrieval is treated as a failure rather than papered over with invented content.

Operational Profile

12.90s and 21.16s for individual calls, 1.44s MCP latency, 212.95s total. Slow for the overall performance shown. 0.004277 USD per run. Cheap on price, but inefficient in time per unit of usable response quality.

Conclusion & Recommendation

Suitable for supervised research and tooling pipelines in which a downstream verifier or rule-based post-processor checks the response against tool outputs. Not suitable for compliance, license verification, multilingual knowledge synthesis, or other paths where the verbal condensation itself is the end product. If you deploy GLM-4.7, treat it as a tool-operating retriever with tight source-binding controls — not as the final authority on substantive claims.

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