Violoop Pairs Its Screen-Reading Device With a Small Decision Model, and the Numbers Get Interesting

Violoop has published four live team runs that pair its computer-use harness with Jev, a small, fast decision model from TypeSafe. The idea, according to the company, is to stop asking a large general model to reason from scratch at every step: the harness narrows the current screen state to a short list of valid actions, and Jev picks the next one. Violoop says that division of labor is especially useful for an AI agent that drives mouse and keyboard input through long, repetitive workflows.

The clearest comparison is a QuickBooks Desktop run where both Jev and Claude Opus 5.5 used the same Violoop harness and each made 123 model decisions. Violoop reports total task time of 978.9 seconds for Opus 5.5 versus 127.3 seconds for Jev, median model-decision time of 7,181 milliseconds versus 390 milliseconds, and estimated model cost of $1.38 versus $0.0027. The company says Jev finished about 7.7 times as fast, with a median decision roughly 18 times faster and an estimated model cost roughly 500 times lower; it notes the cost figures use model list rates and describe model usage only, not the total cost of running the workflow. All 30 of 30 rows were checked during entry rather than in a separate audit afterward.

The other three runs are narrower. On the same Tetris seed, Violoop says Jev had placed 282 pieces and cleared 111 lines at the captured snapshot, against 71 pieces and 26 lines for Opus 5.5, with average decision times of 519 milliseconds versus 2,331 milliseconds; the company notes Jev was also playing with faster gravity, at 0.2 seconds versus 0.4 seconds, and that zero errors were recorded for either model in that particular run. Jev read and ranked 450 comments from Violoop's own Kickstarter discussion, 279 top-level threads and 171 replies, in 30 seconds at about $0.014 in estimated model cost. And it completed a 30-question form in 97.6 seconds, with all 30 answers verified and about $0.002 in estimated model cost.

Violoop is careful about what these runs do and do not prove. The company calls them specific demonstrations rather than an independent benchmark or a promise that every task will show the same performance, and it says each new workflow still needs adaptation at the harness and scheduling layers, so results will vary by application, task shape, and environment. It also says it is researching whether models such as Jev can run on Violoop hardware, and describes that as early research rather than a current local capability or a product commitment.

The device itself is the context for all four runs. Violoop describes a plug-in AI box that reads the screen on the device itself, with an RK3576 octa-core processor and a dedicated AI accelerator rated at 26 TOPS, a Qwen 8B Q4 model running locally at a published 53 tokens per second, 13 GB of memory, 128 GB of storage, HDMI 2.0 capture at 4K@30Hz, dual-monitor support, an on-device AMOLED touchscreen, and USB-HID keyboard and mouse control driven by a dedicated STM32H563 security MCU. The company says nothing executes without a physical key press, that the button wires to the security chip rather than the AI, and that no malware or prompt injection can fake a press.

Violoop is live on Kickstarter, where the company says over US$1 million was pledged on day one and the total now stands at US$1.9 million and counting. It expects shipping to start in late October or early November 2026, with the current Super Early Bird pledge listed at HKD 3,128, about USD 399, and a published retail figure of $699. Violoop says pricing, specifications, shipping timelines and availability may change before final production.

Source: Violoop (https://violoop.ai/blog/violoop-jev-computer-use-live-tests/).