What people are shipping with Jev.

判定特化AI「Jev」に将棋を指させてみた。 Jevは文章を生成せず、選択肢ごとの確率だけを返すモデル。毎手、合法手ぜんぶを渡して1手を選ばせた。 結果は将棋エンジンの最弱・弱に勝ち、中以上には負け。5局で約$0.035。 ロリポップ！AIゲートウェイ経由で今週無料です！ https://t.co/2oni7X6gOD https://t.co/i7urHpN5Kn

@drewpb123 @chesscom Feel free to play with Jev here https://t.co/e36Tx1HOYK Source: https://t.co/jhhVZ89cd0 https://t.co/l85m3q8N4l

@typesafeai @CompleteSkeptic @dabit3 @rileybrown @jackcheng @tamarajtran P.S. I will test Jev hard all week. I will share all learnings here. Build with these models on brewkeg and save 60% on AI model spend. Top up once

@MiaAI_lab Would love to see any results from your findings! As an experiment, could Jev help quickly compact small context windows like: https://t.co/L3O4WPEJMJ

5/ The early usage is already interesting. One developer reported roughly 5,000 requests for about $2 across classification, routing, intent, and steering. A browser agent built with Jev pulled 1.8M views in a day.

Show HN is up if you want the boring production layer around Jev: policy + confidence gate + shadow + recipes + action-level evals Measured on our machine: Claude CLI 48.9s → Jev+harness 1.3s on the same 24-row NL filter