MachinePulse 基于 Gemma 4 微调出 Karpo Decision 31B,多项基准超过 Gemini Decision 1
MachinePulse 拿开源的 Gemma 4 微调出 31B 小模型,When2Call 得分 90.4%,反超 Google 自家 Gemini Decision 1 的 79.9%。单卡 95ms 延迟,权重开放可自部署。做智能体的开发者该看看,工具调用判断层可以不用闭源 API 了。
MachinePulse 将 Google 开源的 Gemma 4 微调为 31B 参数的 Karpo Decision 31B,专攻智能体执行过程中的小型决策:是否调用工具、是否先询问用户、点击页面哪个元素。在 When2Call 基准上,Karpo Decision 31B 得分 90.4%,Gemini Decision 1 为 79.9%;Mind2Web 基准上则是 90.0% 对 86.2%。
该模型在 Decision Index 排行榜上排名第一,该榜单覆盖 300 个模型并使用私有测试集评分,同时 Karpo Decision 31B 在单张 GPU 上的中位延迟为 95 ms。技术路线是拿开源权重做领域微调,而不是从头训练,成本和部署门槛都更低。
对开发者而言,智能体的判断层不再必须依赖闭源 API,权重可以下载后自行部署。与调用 Gemini Decision 1 相比,自托管方案在延迟和成本上更可控,适合对数据隐私和响应速度有要求的智能体项目。
Google's open Gemma 4 just got fine-tuned into a model that beats Google's own Gemini Decision 1.
MachinePulse built Karpo Decision 31B for the small calls an agent makes at every step, like whether to call a tool or ask first, or which element on a page to click. In MachinePulse's runs against Gemini Decision 1: - When2Call (call a tool, ask, or hold back): 90.4% vs 79.9% - Mind2Web (which element to click next): 90.0% vs 86.2%
It also sits at the top of the Decision Index leaderboard (300 models, scored on private tests), with a median latency of 95 ms on a single GPU.
What's interesting is where this leaves agent builders. The judgement layer no longer has to sit behind a closed API. You can download it and run it yourself.
The weights are open, and you can try the model here: https://t.co/66ZodXtNkx
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Karpo Decision 31B, Gemma 4, Gemini, agents.
Source attribution: curated from kimmonismus (@kimmonismus).