亚马逊工程师复刻 Jev 做本地决策模型
Marc Brooker 用 Qwen3.5-2B 加适配器复刻 Jev 的决策输出，单张 RTX 3090 上中位 115ms、零调用成本，权重与脚本已开源。
One Amazon engineer saw Jev become the hottest AI model of September, built his own version that runs on a gaming GPU, and put the weights, the data and the scripts online 16 days later. Jev charges $0.042 per million input tokens. his version costs $0 per call. his name is Marc Brooker. distinguished engineer at Amazon. he saw Jev, tried to build his own take, and it grew into a release with two co-authors. Jev gives no text, just decisions: you give it a question and the answers, it picks one and says how sure it is. Amazon's Strands Decider 2B does the same three answer types, with a median of 115 ms on one RTX 3090. I only noticed this while building the 3D video for this post: the brain inside is not Amazon's. it is Alibaba's Qwen3.5-2B with a small adapter on top. Cloudflare's open Clef, out the same day, is built on Qwen too. 16 days. Sep 15 Jev. Sep 29 OpenAI previews its own Decisions API. Oct 1 Cloudflare and Amazon. the catch is real: 167 of 231 on the JevBench public set, Amazon never claims it beats Jev, and the weights still need your GPU and your power bill. Amazon's own post says it plainly: "You can use, and even train, strands-decider-2b on hardware you already have." most teams paying per token have never checked if a free version runs on a gaming GPU. send this to the one who signs your API bill.
这条是引用着另一条写的：x.com/thegreatest_sv/status/2102802378378621355