相比之下，32GB 内存的 M1 Max 用 llmfit 查一下，最多也就只能跑一跑 2 或 4bit 量化 35b 左右的模型了：
В Иране пообещали заставить США пожалеть о своей агрессии против республики02:08
(who I believe to be a relying party)，这一点在新收录的资料中也有详细论述
When you walk into a room with paying customers, cash flow, and leverage, you’re the pilot — and investors are just along for the ride.。新收录的资料是该领域的重要参考
端侧Agent是这一轮入口战争的核心工具。具体来看，小米将自家的MiclawAgent深植手机底层系统，覆盖手机、电视、汽车等设备；阿里千问整合AI办事入口，实现一句话完成下单和服务请求。巨头通过Agent，不再依赖用户主动打开特定应用，而是让AI自主选择平台和服务完成任务。换句话说，App开始退化为“服务节点”，真正的入口成为执行用户意图的Agent。
Sarvam 105B is optimized for agentic workloads involving tool use, long-horizon reasoning, and environment interaction. This is reflected in strong results on benchmarks designed to approximate real-world workflows. On BrowseComp, the model achieves 49.5, outperforming several competitors on web-search-driven tasks. On Tau2 (avg.), a benchmark measuring long-horizon agentic reasoning and task completion, it achieves 68.3, the highest score among the compared models. These results indicate that the model can effectively plan, retrieve information, and maintain coherent reasoning across extended multi-step interactions.，推荐阅读PDF资料获取更多信息