Computer vision is the field of software that turns images and video into something a system can act on: a box around a forklift, a count of people at a gate, or a sentence explaining what is happening on a loading bay. Two families of vision models dominate practical discussions. Fast, trained detectors such as YOLO. Vision-language models (VLMs) that pair an image encoder with a large language model. Alongside them sits a newer kind of model that does not look at images at all but can sit next to them: TypeSafe AI's Jev, a fast model for typed decisions. Many real systems use more than one.
How YOLO-style detectors work
The original YOLO paper (Redmon et al., 2016) framed detection as a single regression problem: one neural network looks at the whole image once and predicts bounding boxes and class probabilities together. The authors reported real-time speed (45 frames per second for the base model on their test hardware), which is where the "You Only Look Once" name comes from.
The best-known current implementation is from Ultralytics, whose detection documentation describes models pretrained on the COCO dataset, offered in several sizes from nano to extra-large, and exportable to formats such as ONNX, TensorRT, CoreML and LiteRT for different hardware. The trade-off is that a detector only knows the classes it was trained on. If you need it to recognise "a pallet blocking a fire exit", you usually collect and label your own images and fine-tune.
How vision-language models work
VLMs connect pictures to text. CLIP (OpenAI, 2021) learned this by matching about 400 million image-caption pairs from the internet, which lets it classify images against labels written in plain language without task-specific training. Later models feed image features into a language model so they can answer questions, read text in images or describe a scene. Examples include Qwen-VL, Google's PaliGemma, Microsoft's Florence-2 (which handles captioning, detection and grounding through text prompts) and GPT-4V-class hosted models.
A middle ground is open-vocabulary detection. Grounding DINO and Google's OWL-ViT return boxes for objects named in a text prompt, even categories they were never explicitly trained to detect. Accuracy on niche categories still needs testing on your own footage.
The weaknesses are well known. Large VLMs are slower and costlier per image than a small detector, hosted ones send images to a third party, and they can hallucinate: describing objects or events that are not there, or stating a wrong count with confidence.
What Jev is, and what it is not
Jev comes from TypeSafe AI, a company founded by Diogo Almeida, who previously worked at OpenAI. In its launch post the company describes Jev as its first "System One Model", released in early access. The name of the model class refers to Daniel Kahneman's fast, intuitive "System 1" thinking; Jev itself is named after the economist William Stanley Jevons.
Jev takes unstructured state in, such as text or structured program state, and returns type-safe structured decisions. The possible outputs are defined in advance in a schema, and every answer comes with calibrated probabilities and confidence scores. To get there it gives up free-text generation: Jev cannot write sentences or explanations. Instead of producing one token at a time, it outputs all of its answers in parallel in a single query. TypeSafe trains it with a method it calls Reinforcement Learning for Calibrated Decisions (RLCD). The post notes that Jev supports up to 255 options for a single choice.
The company's own figures, which are claims rather than independent results: end-to-end responses of 70 to 500 milliseconds, input priced at $0.042 per million tokens, and output free. On its own workflow evaluations it reports up to 193.6x faster and 444.6x cheaper than large language models, and says itself that these figures are likely on the high end of real-world gains. TypeSafe also says Jev "can't hallucinate". Read that precisely: because outputs must match the schema, Jev cannot return a malformed value or a type error, but it can still pick the wrong answer.
Most importantly for this article, Jev is not a vision model today. TypeSafe's demos run on structured state as text, and the company notes they are "not on images (yet)". Its place in a vision system is after a detector, as a fast decision layer.
Side-by-side comparison
| Factor | YOLO-style detector | VLM / open-vocabulary model | Jev (TypeSafe AI) |
|---|---|---|---|
| Input | Images and video frames | Images plus a text prompt | Text or structured data; not images today |
| Main output | Boxes, classes, confidence scores | Text answers, captions, or boxes from a text prompt | Typed decisions from a predefined schema, with calibrated probabilities; no free text |
| Speed and latency | Built for real time; small models can run on edge devices | Usually slower; large models need GPUs or a cloud API | Company claims 70-500 ms per call; hosted service |
| Labelled data | Needs labelled images for custom classes | Little or none to start; prompts replace labels | Decisions defined in a schema and instructions; test on your own cases |
| Open vocabulary and reasoning | Fixed class list, no reasoning | Strong: free-form questions and explanations | Classify, score, route or branch; cannot explain in words |
| Hallucination risk | Misses and false boxes, but no invented narratives | Can describe things that are not there | No type errors (schema guaranteed), but answers can still be wrong |
| Cost and privacy | Low running cost; can stay fully on-site | Per-call fees if hosted; images may leave your network unless self-hosted | Company lists $0.042 per million input tokens, output free; data goes to a third-party API; early access |
When to use each
- Use a YOLO-style detector when the objects are known in advance (people, vehicles, helmets, parcels), you need results in milliseconds, and cameras or edge boxes must work without the cloud.
- Use a VLM when the question changes often, needs explanation, or involves reading text and context, and when a few seconds per image is acceptable.
- Consider Jev when detections are already turned into structured data and you need fast, typed decisions with a confidence score: classify, route, prioritise or decide whether to escalate. It is in early access, so test it against your own cases before relying on it.
Hybrid pipelines: detection, fast decisions, then explanation
The three can work in sequence. An illustrative example: a YOLO model runs on every frame from site cameras and outputs detections such as "person, zone B, no helmet detected, 02:14". Those detections, as structured data, go to Jev with a fixed schema, for example an action of ignore, log or escalate, and a priority of low, medium or high, each with a probability. Only frames that Jev escalates, or where its confidence is low, go to a VLM with a prompt such as "Is this person wearing a helmet and high-visibility vest? Answer yes or no and explain." The result, plus the snapshot, is passed to an AI agent or workflow that opens a ticket or alerts staff. VLM cost and data exposure stay limited to the frames that matter. Adding a RAG step can let the VLM check site rules before it answers.
Keep a human in the loop for consequential decisions, log what each model saw and said, and treat cameras as part of your IoT security estate. Connected devices are an attack surface, as our article on AI worms explains. Privacy rules on recording people also apply; check with a legal adviser where needed.
technine.io designs camera, sensor and alert workflows as part of our AIoT work, and broader model integrations through AI system development services.
Frequently asked questions
Is YOLO outdated now that VLMs exist?
No. For fixed, known objects at high frame rates, a small trained detector is usually faster, cheaper and easier to run on-site. VLMs add flexibility and explanation, not raw throughput.
Can a VLM replace labelling my own data?
For prototypes, often yes: open-vocabulary models and prompts can get you started. For production accuracy on your specific cameras and objects, you will usually still need a labelled test set, and often training data.
Can Jev analyse camera images directly?
Not today. TypeSafe says its demos run on structured text state and not on images yet. In a vision system, Jev would act on detector output or other structured data, and it cannot write explanations, so a VLM or a person still handles cases that need one.
How do I reduce VLM hallucination in a vision system?
Ask narrow, structured questions, send only the relevant frame or crop, check answers against detector output or rules, and route uncertain cases to a person.