AutoArk团队(Edge0 Team)开源了全球首个拥有1450亿参数的端侧AI模型Edge0-145B及配套统一推理框架 1。该模型旨在让高性能编码、推理和Agent能力在个人电脑上本地运行 1。通过架构、训练、加载和调度四层设计创新,Edge0-145B在保持仅约100亿参数参与单次计算的同时,实现了媲美云端前沿模型的性能,显著降低了内存占用并消除了云API成本 1。
在性能指标方面,Edge0-145B在笔记本电脑上可实现80+ tokens/s的生成速度,峰值内存为28GB 1。其前代模型Edge0-35B在iPhone 18 Pro Max上运行时达到37+ tokens/s的速度,且内存低于2.5GB 1。基准测试得分显示,该模型在DeepSWE 1.1中得分为64.6,优于Qwen3.8-Flash-Next的58.7分;GPQA-Diamond得分为90.4;IFBench得分为80.6;Terminal-Bench 2.1得分为86.52 1。
核心技术创新包括稀疏专家共享矩阵架构、面向本地执行的训练策略、GPU/系统内存/SSD三级流水线加载机制以及CPU与GPU协同调度 1。应用案例显示,Edge0-145B能在本地完成复杂的视频制作、3D建模(如曼哈顿城市场景)、游戏开发及硬件控制(如无人机编程),无需云端Token费用 1。支持平台包括macOS、Windows、Linux,兼容Apple M系列芯片及NVIDIA RTX Spark、4090、5090等硬件 1。
AutoArk’s Edge0 Team has open-sourced the world’s first edge AI model with 145 billion parameters, accompanied by a unified inference framework designed to bring high-performance coding, reasoning, and agent capabilities directly to personal computers 1. This initiative aims to enable local execution of tasks previously requiring cloud infrastructure, significantly reducing memory usage and eliminating cloud API costs 1. The model features a unique architecture where, despite its total size of 145B parameters, only approximately 10B parameters are actively computed during each inference step 1.
In terms of performance, Edge0-145B achieves a generation speed of over 80 tokens per second on laptops while maintaining a peak memory footprint of 28GB 1. Its predecessor, the Edge0-35B model, demonstrated efficiency on mobile devices by reaching speeds above 37 tokens per second on an iPhone 18 Pro Max with less than 2.5GB of memory usage 1. Benchmark results highlight the model's competitive edge, scoring 64.6 on DeepSWE 1.1 (surpassing Qwen3.8-Flash-Next’s 58.7), 90.4 on GPQA-Diamond, 80.6 on IFBench, and 86.52 on Terminal-Bench 2.1 1.
The technical foundation relies on four layers of innovation: a sparse expert shared matrix architecture, training strategies optimized for local execution, a three-tier pipeline loading mechanism utilizing GPU, system memory, and SSD, and coordinated CPU-GPU scheduling 1. These advancements support complex local applications such as video production, 3D modeling (including Manhattan cityscapes), game development, and hardware control like drone programming, all without incurring token fees from cloud services 1. The models are compatible with macOS, Windows, and Linux platforms, supporting hardware ranging from Apple M-series chips to NVIDIA RTX Spark, 4090, and 5090 GPUs 1.
评论
还没有评论,欢迎留下第一条。