Jev in Practice: Browser Use and Agent Workflows

The author ran Jev with Cua Driver on 9 executable tasks and it passed 8, versus 9 of 9 for Astra/Codex, with median decisions 13.6x faster and median task time 27% lower.

The author demonstrates Jev generating images by predicting each pixel in parallel, showing how it can be used for image generation tasks.

A small crypto trading demo has Jev answer buy or wait on real BTC, ETH and XRP prices, returning inspectable JSON and executing no trades.

A diagram explains how Jev uses a single Transformer decoder pass with cached keys and values to pick each JSON field's token directly, replacing token-by-token generation.

A Jev beta tester uses it as a router that decides whether a chunk of text goes through deterministic rules or a more expensive model, splitting work by cost.

The author breaks character-by-character generation into 29 yes/no classification questions, using Jev as the classifier in an autoregressive loop that produces a text-continuation model.

The author demonstrates making the Jev decision model conversational, with a video showing the interaction.

A developer backtests Jev against a Gemini-2.5-flash-lite setup in their MCP-based multiuser spec governance system, finding it cheaper and more consistent but slower than expected.

After getting access, the author launched a web page called almost-certain to experiment with Jev's generation capabilities and shared the link.

The author used Typesafe's new Jev model to play Super Mario Bros., citing fast inference and structured outputs as making it surprisingly effective for real-time use cases.

The author built a codebase classifier with Jev as a possible fix for overengineered code produced by agents, and asked for suggestions on what to test next.

After about 5,000 Jev requests costing roughly $2, the developer says Jev is far faster and cheaper than LLM classifiers for classification, model routing, intent detection and steering.