Edge0开源145B端侧大模型，本地运行仅需29GB显存
Edge0-145B: Frontier AI Capabilities on Your Own Computer
端侧大模型突破百B参数大关是重要进展，提供了完整的本地部署方案与性能数据，开发者可参考其架构设计优化推理效率。
Today, we're open-sourcing Edge0-145B and Edge0-120B, together with our unified inference framework and deployment tools. Edge0-145B is the world's first open-source on-device AI model with 145 billion parameters. It brings coding, reasoning, and agent capabilities that rival frontier cloud models to your own computer.
今天，我们开源了 Edge0-145B 和 Edge0-120B，以及我们的统一推理框架和部署工具。Edge0-145B 是全球首个拥有 1450 亿参数的开源端侧 AI 模型。它将媲美前沿云端模型的编码、推理和智能体能力带到了你自己的电脑上。
On-device AI has long forced a tradeoff between model capability and what the hardware can run. Larger models demand more memory, bandwidth, and power; smaller models can struggle with complex work. Moving that work to the cloud adds model-provider fees, data-handling constraints, and a dependence on connectivity.
端侧 AI 长期以来一直迫使用户在模型能力和硬件运行能力之间做出权衡。更大的模型需要更多的内存、带宽和电力；而较小的模型在处理复杂任务时可能会力不从心。将工作转移到云端则带来了模型提供商的费用、数据处理限制以及对网络连接的依赖。
This release addresses those constraints in four ways:
本次发布通过以下四种方式解决了这些限制：
| The constraint | What Edge0-145B makes possible |
|---|---|
| Large models are difficult to run on personal hardware. | A 145B model runs locally on a compatible laptop, with reported peak memory of 29GB and decoding above 80 tokens/s. |
| Small on-device models can fall short on demanding tasks. | Coding, reasoning, and instruction-following results are competitive with specific frontier cloud models, as the comparisons below show. |
| A useful assistant needs to finish work, not just answer questions. | The model plans, writes and runs code, debugs errors, and checks results visually for projects such as 3D scenes, games, hardware programs, and PCB designs. |
| Cloud workflows bring recurring fees, sensitive-data restrictions, and network dependence. | Local inference incurs no corresponding cloud-model token or API fees; data can stay on the computer, and core capabilities keep working offline. Hardware and electricity still cost money. |
| 限制 | Edge0-145B 实现的可能 |
|---|---|
| 大型模型难以在个人硬件上运行。 | 一个 145B 的模型可以在兼容的笔记本电脑上本地运行，据报道峰值内存为 29GB，解码速度超过每秒 80 个 token。 |
| 小型端侧模型在 demanding 任务中可能表现不足。 | 编码、推理和指令遵循的结果与特定的前沿云端模型具有竞争力，如下面的比较所示。 |
| 有用的助手需要完成工作，而不仅仅是回答问题。 | 该模型能够规划、编写和运行代码，调试错误，并针对 3D 场景、游戏、硬件程序和 PCB 设计等项目进行视觉结果检查。 |
| 云端工作流程带来 recurring 费用、敏感数据限制和网络依赖。 | 本地推理不会产生相应的云端模型 token 或 API 费用；数据可以保留在计算机上，核心功能在离线状态下也能继续工作。硬件和电力仍然需要成本。 |
We're releasing the full stack so developers can reproduce the results, evaluate or fine-tune the models, and integrate them into their own products.
我们正在发布完整的软件栈，以便开发人员可以复现结果、评估或微调模型，并将其集成到他们自己的产品中。
Get the model weights, explore the inference framework, or try the Edge0 App.
获取模型权重，探索推理框架，或试用 Edge0 App。
How We Make 100B+ Parameter AI Run on a Personal Computer
如何让 1000 亿+ 参数的 AI 在个人电脑上运行
Edge0-145B Laptop Performance: Memory, Speed, and Power
Edge0-145B 笔记本性能：内存、速度和功耗
Edge0-145B makes a 145-billion-parameter model usable on a personal laptop. A useful local model has to fit in memory, process prompts promptly, generate fast enough to use, and keep power consumption under control. It also needs to sustain longer tasks.
Edge0-145B 使 1450 亿参数的模型能够在个人笔记本电脑上使用。一个有用的本地模型必须能够装入内存，及时提示处理，生成速度快到足以使用，并将功耗控制在合理范围内。它还需要能够支持更长的任务。
| Requirement | Reported Edge0-145B result | What it enables |
|---|---|---|
| Memory | 29GB peak memory | Runs on a compatible personal laptop, such as a MacBook Pro, without a server. |
| Prompt processing | 1,100+ tokens/s prefill | Processes long prompts, including dozens of pages of code, before generation begins. |
| Generation | 80+ tokens/s decode | Generates faster than a person reads. |
| Power | Approximately 120W typical power | Sustains operation without throttling in the reported setup. |
| Long tasks | 128K–256K context | Maintains essentially the same speed on long inputs, supporting extended work. |
| 需求 | 报告的 Edge0-145B 结果 | 它带来的能力 |
|---|---|---|
| 内存 | 峰值内存 29GB | 可在兼容的个人笔记本电脑（如 MacBook Pro）上运行，无需服务器。 |
| 提示处理 | 预填充速度 1,100+ tokens/s | 在生成开始前处理长提示，包括数十页代码。 |
| 生成 | 解码速度 80+ tokens/s | 生成速度快于人类阅读速度。 |
| 功耗 | 典型功耗约 120W | 在报告的设置中可持续运行而不降频。 |
| 长任务 | 128K–256K 上下文 | 在处理长输入时保持基本相同的速度，支持扩展工作。 |
Our earlier Edge0-35B phone release reported peak memory below 2.5GB. On an iPhone 18 Pro Max, Edge0-35B delivers 37+ tokens/s using approximately 2.5GB of memory. Edge0-145B has roughly four times as many parameters; its laptop comparison reports 80+ tokens/s and 28GB peak memory. These measurements apply to different models and hardware.
我们此前发布的 Edge0-35B 手机版本报告峰值内存低于 2.5GB。在 iPhone 18 Pro Max 上，Edge0-35B 使用约 2.5GB 内存可实现 37+ tokens/s 的速度。Edge0-145B 的参数数量约为前者的四倍；其笔记本对比报告显示速度为 80+ tokens/s，峰值内存为 28GB。这些测量值适用于不同的模型和硬件。
On-device model scale over time. The chart covers 67 open models of at least 1B parameters from 14 labs, intended for phones, laptops, and AI PCs. Same-size base, instruct, and reasoning variants count once; each line tracks a lab's largest model, including releases before July 2025. Filled points mark new highs and hollow points other releases. The scale is logarithmic below 40B and linear above it. Apple is excluded because its weights are not released. Sources: vendor announcements and model cards.
端侧模型规模随时间的变化。该图表涵盖了来自 14 家实验室的至少 10 亿参数的 67 个开源模型，旨在用于手机、笔记本电脑和 AI PC。相同大小的基础、指令和推理变体仅计一次；每条线追踪各实验室的最大模型，包括 2025 年 7 月之前发布的版本。实心点标记新高，空心点表示其他发布版本。400 亿参数以下采用对数刻度，以上采用线性刻度。苹果被排除在外，因为其权重未公开。来源：厂商公告和模型卡片。
Why On-Device Models Have Stayed Small
为什么端侧模型一直很小
Memory limits kept most on-device models below 10B parameters before September 2026. Apple Intelligence's on-device foundation model had approximately 3B parameters, while Google's Gemini Nano came in 1.8B and 3.25B versions. Even a 7B model could consume approximately 4GB of memory.
在 2026 年 9 月之前，内存限制使大多数端侧模型的参数量保持在 100 亿以下。Apple Intelligence 的端侧基础模型约有 30 亿参数，而 Google 的 Gemini Nano 有 18 亿和 32.5 亿两个版本。即使是一个 70 亿参数的模型也可能消耗约 4GB 内存。
Larger models faced a memory wall. Model weights occupy storage and must be read during generation, competing for limited memory and bandwidth with the operating system and other apps. Even an aggressively compressed 35B model could need approximately 20GB of memory, far more than a phone could make available.
更大的模型面临内存墙。模型权重占用存储空间，并在生成过程中必须被读取，与操作系统和其他应用争夺有限的内存和带宽。即使是经过激进压缩的 350 亿参数模型，也可能需要约 20GB 内存，远超手机可提供的容量。
Compute, bandwidth, and power created competing constraints. Larger models offered more capability, but strained memory capacity, generation speed, and battery life. Making models smaller helped them fit, often at the cost of capability.
计算能力、带宽和功耗形成了相互制约的因素。更大的模型提供了更强的能力，但会加剧内存容量、生成速度和电池寿命的压力。缩小模型尺寸有助于其适配设备，但往往以牺牲能力为代价。
Privacy alone could not close that gap. Apple Intelligence protected user data, but Apple acknowledged that its on-device model was not designed for agentic applications requiring general world knowledge. The prevailing assumption was that on-device intelligence could not match the cloud.
仅靠隐私保护无法弥补这一差距。Apple Intelligence 保护了用户数据，但苹果承认其端侧模型并非为需要通用世界知识的智能体应用而设计。 prevailing assumption（普遍假设）是端侧智能无法与云端相媲美。
Under the Hood of Edge0: Four Layers of Model and Inference Design
Edge0 的底层架构：模型与推理设计的四个层次
We designed Edge0-145B and its inference engine together to reduce the memory and data movement that make large models difficult to run locally. The design spans architecture, training, loading, and scheduling.
我们共同设计了 Edge0-145B 及其推理引擎，以减少使大型模型难以在本地运行的内存占用和数据移动。该设计涵盖了架构、训练、加载和调度。
Architecture: share large matrices across sparse experts. Sharing matrices substantially reduces the weights that must be stored and moved while keeping the expert count unchanged. Each group of 64 experts shares two large matrices, while each expert retains a small set of parameters that captures its differences. Previously, every expert stored a complete set of matrices. Of the model's 145B total parameters, only approximately 10B participate in computation on each inference step.
架构：在稀疏专家间共享大型矩阵。共享矩阵大幅减少了必须存储和移动的参数权重，同时保持专家数量不变。每 64 个专家组共享两个大型矩阵，而每个专家保留一组小型参数以捕捉其差异。此前，每个专家都存储完整的矩阵集。在模型总共 1450 亿参数中，每次推理步骤仅有约 100 亿参数参与计算。
Training: build for local execution from day one. We train the model in its on-device architecture from the outset, with deployment constraints built into its design. This avoids treating local deployment as a compression or pruning step after training a cloud model.
训练：从一开始就为本地执行而构建。我们从一开始就在其端侧架构下训练模型，并将部署约束融入设计中。这避免了将本地部署视为云模型训练后的压缩或剪枝步骤。
Loading: pipeline GPU memory, system memory, and SSD access. Reading and computation overlap across three storage tiers, avoiding serial waits. Frequently used “hot experts” remain in GPU memory. An expert library of approximately 40GB sits in system memory, with experts brought into GPU memory as needed. Approximately 48GB of vocabulary data resides on the SSD and is read on demand.
加载：流水线化 GPU 内存、系统内存和 SSD 访问。读取与计算在三个存储层级间重叠进行，避免串行等待。常用的“热专家”保留在 GPU 内存中。约 40GB 的专家库位于系统内存中，根据需要调入 GPU 内存。约 48GB 的词表数据存储在 SSD 上，按需读取。
Scheduling: use the CPU and GPU together. The GPU computes cached experts while the CPU handles the remaining work in parallel. Speculative decoding verifies multiple candidate tokens at once, spreading the cost of reading weights across more output tokens. These choices support the reported 80+ tokens/s decode rate.
调度：CPU 与 GPU 协同工作。GPU 计算缓存的专家，而 CPU 并行处理剩余工作。推测解码一次性验证多个候选 token，将读取权重的成本分摊到更多的输出 token 上。这些选择支持报告中的每秒 80+ token 的解码速率。
Model and inference-engine design work together to make 100B+ parameter inference practical on personal computers. Source: internal testing.
模型与推理引擎的设计协同工作，使得在个人电脑上实现 1000 亿以上参数的推理成为可能。来源：内部测试。
Together, these changes let parameter counts grow from tens of billions to more than a hundred billion without proportional increases in memory use, power consumption, and latency.
综上所述，这些变化使得参数量从数百亿增长到一千多亿，而内存使用、功耗和延迟并未成比例增加。
What Edge0 Can Do: Anything From Chat to Complex Work on Your Computer
Edge0 能做什么：从聊天到电脑上的复杂任务
Edge0-145B can carry a complex task through to a usable result. Its coding and reasoning capabilities support local workflows that combine planning, tool use, execution, debugging, and visual checks. The benchmarks establish its capabilities; the demonstrations show how it puts them to work.
Edge0-145B 能够将复杂任务执行至产生可用结果。其编码和推理能力支持本地工作流，该工作流融合了规划、工具使用、执行、调试和视觉检查。基准测试确立了其能力；演示则展示了如何将其付诸实践。
Coding and reasoning that rival frontier cloud models
编码与推理能力媲美前沿云端模型
Edge0-145B is competitive with specific frontier cloud models across coding, reasoning, knowledge, and instruction following:
在编码、推理、知识和指令遵循方面，Edge0-145B 与特定前沿云端模型具有竞争力：
- Coding: It scores 64.6 on DeepSWE 1.1, ahead of Qwen3.8-Flash-Next at 58.7 and DeepSeek-V4-Flash-0731 at 54.4.
- Reasoning and knowledge: It scores 90.4 on GPQA-Diamond, close to DeepSeek-V4-Flash-0731 at 90.8 and Claude-Opus-4.6 \(Max\) at 91.3.
- Instruction following: It scores 80.6 on IFBench, ahead of DeepSeek-V4-Flash-0731 at 79.2 and Claude-Opus-4.6 \(Max\) at 62.5.
- 编码：它在 DeepSWE 1.1 上得分 64.6，领先于 Qwen3.8-Flash-Next（58.7）和 DeepSeek-V4-Flash-0731（54.4）。
- 推理与知识：它在 GPQA-Diamond 上得分 90.4，接近 DeepSeek-V4-Flash-0731 的 90.8 和 Claude-Opus-4.6 \(Max\) 的 91.3。
- 指令遵循：它在 IFBench 上得分 80.6，领先于 DeepSeek-V4-Flash-0731 的 79.2 和 Claude-Opus-4.6 \(Max\) 的 62.5。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力