Camera Intelligence — October 2026: Warehouse Safety & Everyday Work
OpenAI gpt-6-luna · TypeSafe AI Jev 1.14.0 · LiquidAI d1-2B (bfloat16) · Qwen3.5-9B (Q4_K_M)
- Written by
- Agentinosaurus Team
- Published
- Reading time
- 4 min read
Moving loads, stationary forklifts and pedestrian walkways. Five factory recordings show why the details of an Event matter.
A specific rule beats a vague warning
What should a camera notice in a working warehouse? For this comparison, we asked two concrete questions: is a moving forklift carrying more than two containers, and is someone walking outside the green pedestrian walkway for more than five seconds?
Five short factory recordings show moving loads, a stationary forklift, machine work and pedestrian movement. Every model received the same images and the same two Events.
A moving forklift with a larger stack
The forklift drives across the scene carrying more than two orange containers. All four models matched the load Event.
| Model | Label agreement¹ | False alerts² | Median response | API cost | RAM³ GiB | VRAM³ GiB |
|---|---|---|---|---|---|---|
| OpenAI gpt-6-luna | 100% | 0% | 1.86s | $0.00059 | — | — |
| TypeSafe AI Jev 1.14.0 | 100% | 0% | 0.82s | $0.00061 | — | — |
| LiquidAI d1-2B (bfloat16) | 100% | 0% | 2.55s | — | 4.27 | 5.18 |
| Qwen3.5-9B (Q4_K_M) | 100% | 0% | 2.54s | — | 7.38 | 6.65 |
Working beside a machine
The worker stays near the machine and makes short movements. That is different from continuously walking outside the walkway for more than five seconds. All four models stayed quiet for both Events.
| Model | Label agreement¹ | False alerts² | Median response | API cost | RAM³ GiB | VRAM³ GiB |
|---|---|---|---|---|---|---|
| OpenAI gpt-6-luna | 100% | 0% | 1.24s | $0.00043 | — | — |
| TypeSafe AI Jev 1.14.0 | 100% | 0% | 0.80s | $0.00044 | — | — |
| LiquidAI d1-2B (bfloat16) | 100% | 0% | 2.39s | — | 4.27 | 5.18 |
| Qwen3.5-9B (Q4_K_M) | 100% | 0% | 2.36s | — | 7.38 | 6.65 |
A loaded forklift that stays still
A load is visible, but the Event requires a moving forklift. OpenAI and Qwen returned a match; TypeSafe and Liquid did not. Recognising the objects is only part of the task—the condition matters too.
| Model | Label agreement¹ | False alerts² | Median response | API cost | RAM³ GiB | VRAM³ GiB |
|---|---|---|---|---|---|---|
| OpenAI gpt-6-luna | 50% | 50% | 1.17s | $0.00064 | — | — |
| TypeSafe AI Jev 1.14.0 | 100% | 0% | 0.87s | $0.00067 | — | — |
| LiquidAI d1-2B (bfloat16) | 100% | 0% | 0.66s | — | 4.27 | 5.18 |
| Qwen3.5-9B (Q4_K_M) | 50% | 50% | 2.75s | — | 7.38 | 6.65 |
Walking beside the green walkway
The same person progresses along the grey aisle beside the marked pedestrian route for more than five seconds. Only TypeSafe matched the walking Event. Liquid also returned unknown for the separate load Event in this clip.
| Model | Label agreement¹ | False alerts² | Median response | API cost | RAM³ GiB | VRAM³ GiB |
|---|---|---|---|---|---|---|
| OpenAI gpt-6-luna | 50% | 0% | 1.03s | $0.00059 | — | — |
| TypeSafe AI Jev 1.14.0 | 100% | 0% | 0.67s | $0.00061 | — | — |
| LiquidAI d1-2B (bfloat16) | 0% | 0% | 2.53s | — | 4.27 | 5.18 |
| Qwen3.5-9B (Q4_K_M) | 50% | 0% | 2.74s | — | 7.38 | 6.65 |
Two containers on a moving forklift
The forklift moves with two orange containers, so the more-than-two load Event should not match. OpenAI correctly returned no_match. TypeSafe, Liquid and Qwen returned match: false positives for this Event. All four correctly rejected the separate walking Event.
| Model | Label agreement¹ | False alerts² | Median response | API cost | RAM³ GiB | VRAM³ GiB |
|---|---|---|---|---|---|---|
| OpenAI gpt-6-luna | 100% | 0% | 0.78s | $0.00038 | — | — |
| TypeSafe AI Jev 1.14.0 | 50% | 50% | 0.52s | $0.00039 | — | — |
| LiquidAI d1-2B (bfloat16) | 50% | 50% | 2.24s | — | 4.27 | 5.18 |
| Qwen3.5-9B (Q4_K_M) | 50% | 50% | 2.13s | — | 7.38 | 6.65 |
What this small comparison showed
TypeSafe detected both positive Events, but also falsely alerted on the two-container load. OpenAI and Qwen both alerted on the stationary forklift. Only TypeSafe recognised the qualifying walk outside the green route.
Across the two cases where an Event was present, detection was OpenAI gpt-6-luna: 50%; TypeSafe AI Jev 1.14.0: 100%; LiquidAI d1-2B (bfloat16): 50%; Qwen3.5-9B (Q4_K_M): 50%. Across the eight negatives, false-alert rates were OpenAI gpt-6-luna: 12.5%; TypeSafe AI Jev 1.14.0: 12.5%; LiquidAI d1-2B (bfloat16): 12.5%; Qwen3.5-9B (Q4_K_M): 25%. Liquid also returned one unknown on a negative case.
These five recordings come from one facility. The results illustrate these specific situations, rather than general warehouse accuracy.
Correction to the container count
An earlier edition described the final forklift as carrying three containers. It carries two. The corrected results show that OpenAI stayed quiet correctly, while TypeSafe, Liquid and Qwen raised false alerts.
The exact Events
A moving forklift carries more than two distinct crates, boxes, packages or containers.
A person walks continuously outside the green pedestrian walkway for more than five seconds, without operating a machine or performing another work task.
How to read the tables
¹ Label agreement means the share of the two Events answered correctly in each scene. Unknown is not counted as correct.
² False alerts count matches where the requested Event was absent. The summary’s detection percentage covers only cases where an Event was present.
³ Memory figures show peak usage on an RTX 4080 Laptop GPU, including model loading. Liquid and Qwen use different GPU memory measurements, so these figures are approximate comparisons, not minimum hardware requirements.
Response times measure the model response, not the time from an incident to an alert. Cloud costs are estimates; local hardware and electricity are excluded. “—” means unavailable or not applicable.
Images: Önal & Dandıl (2024), Mendeley Data, CC BY 4.0. Stills have been resized.