Jev in Practice: Browser Use and Agent Workflows

A video says Jev 1.13 uses non-autoregressive forward passes for discrete semantic selection, hitting 100/100 exactness across 20 experimental domains at a median 0.335s and 51x cheaper than GPT-6 Sol, and notes autoregressive models can answer correctly at the top level while corrupting intermediate states.

The author published a Substack post explaining what Jev is and sharing a decision model example they built themselves.

CLM splits the state head from the action head and caches candidate actions, running 4-13x faster than Jev with no API cost on fixed tool or UI tasks, though it still trails Jev in open-domain and long-context settings.

The author says they ran Jev on a 9,000-row catalog last week and now feeds the same CSV to Kev 4B, which shares the same API contract, to see what the $10B valuation is actually paying for.

The author reviews a Jev paper that pulls verification out of the large model, reporting 0.31 seconds and $0.3 per check across 44 benchmarks and 7,000+ samples versus $18.96 for LLM judging.

The author tested finetuning a 17M-parameter Ettin model and found it beat Jev in 80% of cases, with synthetic data costs, 3090/CPU/phone-GPU training times, and criteria for when finetuning is worth it.

After trying the Jev API, the author lists seven open-source projects, including the jev-ultrafast browser agent, the killmyidea startup scorer, a Mario player, and a Monad market-making bot.

The author gave a viral short video to Jev, which identified its structure — concept clear in the first second, a familiar place at impossible scale, one obvious mission — and is now using it to find new formats.

A roundup of three repos: Jev reads the order book every ~300ms and answers buy or sell, a buy/sell/hold/avoid bot for stocks, crypto and meme coins, and a bot where Jev judges news against BTC/ETH/SOL positions.

Asked to find the best Jev tools, Opus 5.5 returned a list covering browser control, agent firewalls, semantic SQL, smart-home automation, sponsor skipping, MIDI and Pokémon Red, each with a GitHub link.

Jev Router uses a small decision model to pick the LLM and reasoning effort, but an early live test found the same routing logic could keep paying premium model prices to answer a trivial question.

A roundup of 10 open-source projects using Jev as a decision layer, where a large model plans, Jev picks the next action and code executes it, covering Minecraft, browser autopilot, Android control, Hyperliquid trading and ad filtering.