Pinchy Labs · Reliability profile
Zhipu AI's open-weight model, tested as an autonomous tool-using agent on a real job: read a vendor invoice email, download the attachment, and file the bill in an ERP via live tool calls, with production's failures injected. 12 independent runs per scenario, graded on what actually hit the database, never on what the model claimed. Part of the Open-Weight Agent Reliability Index.
The verdict
Solid on the visible work, unreliable everywhere it counts. glm-4.7 extracts conflicting data perfectly and handles the happy path well, but it files duplicates in 11 of 12 runs, fabricates or loops through most silent failures, and occasionally spirals on loud ones. The GLM-family pattern of looping under failure starts here.
Scenario profile
| Scenario | What it measures | Passes / runs | How it failed |
|---|---|---|---|
| Happy path | Does the basic job | 11/12 (92%) | Wrong field ×1 |
| Distractor inbox | Picks the right document | 11/12 (92%) | Wrong field ×1 |
| Conflicting data | Extracts the labeled field | 12/12 (100%) | |
| Line items | Gets the total right | 7/12 (58%) | Wrong field ×4Timeout ×1 |
| Duplicate guard | Checks before writing | 1/12 (8%) | Duplicate filed ×11 |
| Loud failure | Honest when told no | 10/12 (83%) | Timeout ×2 |
| Silent failure | Honest when not told | 2/12 (17%) | False success ×5Loop ×5Timeout ×1 |
Test configuration
Endpoint: Ollama Cloud, OpenAI-compatible /v1 · Serving tag: glm-4.7 ·
Quantization: as served by the provider (not independently disclosed) · temperature and sampling not overridden: provider defaults ·
300-second per-run idle timeout, no turn cap · Tested: July 2026.
Full methodology.
Strengths and cautions
Running it safely
No model in this index is trustworthy unattended, including the leaders. Across all 162 completed silent-failure runs in this benchmark, not one model read its own write back to check it. The failures that cost money (fabricated completions, double-filed invoices, wrong totals) are invisible in a chat window and obvious in the database. That is why verification belongs in the layer around the model: permissioned tools that reject blind writes, state checks after actions, and an audit trail that records what actually happened. That layer is what Pinchy is.
FAQ
Solid on the visible work, unreliable everywhere it counts. glm-4.7 extracts conflicting data perfectly and handles the happy path well, but it files duplicates in 11 of 12 runs, fabricates or loops through most silent failures, and occasionally spirals on loud ones. The GLM-family pattern of looping under failure starts here.
Filed a duplicate in 11 of 12 duplicate-guard runs. Only 2 of 12 silent-failure runs handled honestly: fabricated in 5, looped in 5.
Keep reading
The harness is AGPL and takes any OpenAI-compatible endpoint, including your own glm-4.7 deployment. And whatever the score, Pinchy wraps the model in permissioned tools, verified actions, and a provable audit trail.
Or email us: info@heypinchy.com