Start with what you need to know and a practical first project. If you need more detail, you can also review the current release and its checks, how models score feature by feature, and what we confirm before deployment.
The same product should make sense to the person funding it, the team using it, and the people responsible for operating it safely.
Is this a serious product worth evaluating?
Start with what Zy-AI covers, what is available now, and the booking flow. You do not need the technical details first.
Review the productWill this fit the way our people work?
Review the example workflow, setup responsibilities, available support, and what must be agreed before a pilot.
Review implementationWhat technical detail can we review?
Review the security overview, the release checks, and how each model scores when deeper evaluation begins.
Review security and controlsYou do not need a company-wide transformation plan. Start with one recurring task that costs time, creates confusion, or depends too heavily on one person.
Help your team find an approved answer without searching across folders, messages, and old documents.
Prepare what a reviewer needs, with its sources, and keep the decision and any approved change on record.
Ask a connected database in plain words and get the figures worked out the same way each time.
These counts come from the v2.3.2 release record. They show the checks completed by Zyqlo and are not an independent certification.
An AI model is the engine that writes the answers, and Zy-AI works with the one you choose, through your own provider account. We check Zy-AI on open models, whose files are published so a team can run them at low cost, because answers you can check should not need the most expensive model. Zy-AI adds the grounding, checks, and controls around the model. The model still matters, so we show how each one scores, feature by feature. Every answer is graded by Claude Opus, a separate model acting as the judge. The table shows the latest full run of the check, on Zy-AI v1.12.8.
| AI model (test harness) | Runs | Average score | Median seconds per answer | Measured on |
|---|---|---|---|---|
| Qwen 3 30B open | 3 | 0.83 | 18.7s | v1.12.8, Oct 26, 2026 |
| GLM 4.7 Flash open | 1 | 0.76 | 37.2s | v1.12.8, Oct 26, 2026 |
Where they differ most: Qwen on web research (0.70 against 0.20), knowledge sources, following a conversation, and skills; GLM on artifact generation (1.00 against 0.74) and code understanding. They are level on memory, tools, visuals, and richness.
| Feature | Qwen 3 30B | GLM 4.7 Flash |
|---|---|---|
| Your documents and data | ||
| Knowledge sourcesAnswers from your documents | 0.97 | 0.68 |
| DatasourcesQuestions about connected databases | 0.67 | 0.72 |
| AttachmentsFiles and screenshots you share | 0.62 | 0.50 |
| Code understandingExplaining a connected code project | 0.79 | 0.89 |
| Following the conversation | ||
| Conversation contextTopic switches, recaps, and picking up a decision | 0.94 | 0.72 |
| RecallBringing back what you said or asked it to remember | 1.00 | 0.88 |
| MemoryKeeping facts you gave it, and your corrections | 0.94 | 0.94 |
| Getting work done | ||
| ToolsRunning code and searches, without claiming apps it lacks | 1.00 | 1.00 |
| Artifact generationMaking a file you can download | 0.74 | 1.00 |
| Web researchCurrent facts and costs from the web | 0.70 | 0.20 |
| SkillsUsing the right built-in skill, or none | 0.58 | 0.38 |
| VisualsCharts from your figures | 1.00 | 1.00 |
| Answer quality and limits | ||
| RichnessComplete explanations, advice, and drafts | 0.75 | 0.76 |
| GuardrailsSaying plainly what falls outside its rules or sources | 0.97 | 1.00 |
Both models were measured on Zy-AI v1.12.8 on October 26, 2026, against the same check: 14 features covered by 32 scripted conversations run live against Zy-AI. Each answer is scored from 0 to 1 against a written rubric by Claude Opus (the judge model, not a model under test); a feature's score averages its conversations, and the average score averages the 14 features. Qwen 3 30B, the model Zy-AI releases are checked on, ran three times. GLM 4.7 Flash rests on a single run, judged a few hours later rather than in the same sitting, with images read by Llama 4 Scout whichever model answers. A model is listed once it has completed the whole check, with every feature judged. Scores move from run to run by 0.05 to 0.30 depending on the feature, so differences that small are a tie, not a ranking. Answer times are the median across the check. These are Zyqlo's own checks, not an independent certification.
Bring one business question, review step, or report. We will show Zy-AI working on it with safe demonstration data and discuss what a practical first setup would need.
Every customer setup is planned around its intended use, team, and operating environment.
During discovery, we confirm expected capacity, sign-in, storage, backups, recovery, support, and responsibility for each provider.