Skip to content

PhoneClaw v1.3.2 — download bug fix + MTP toggle

Choose a tag to compare

View all tags
@kellyvv kellyvv released this 09 May 05:34
· 202 commits to main since this release
v1.3.2
4ebaf39

PhoneClaw v1.3.2 — local AI agent for phones

🧠 默认端侧运行 Gemma 4 大语言模型，支持文本 / 图片 / 语音多模态对话；联网仅用于用户显式触发的功能

🌐 English version below — scroll down for English release notes

🆕 本次更新 (v1.3.2)

  • 🐛 修复模型下载完成后又自动重新下载的 bug：之前 downloader 用写死的 expectedFileSize 跟 HF/HF-mirror 实际文件大小做硬校验，差几 MB 就把刚下完整的文件误删并触发重下。新版改为只信任 HTTP 服务器返回的 Content-Length / Content-Range，不再用常量做精确字节级校验。同步修了暂停后续传时的同病双胞胎路径。
  • ⚡ 新增 MTP 推测解码开关（实验性）：在配置页可开启 Gemma 4 的 Multi-Token Prediction 推测解码。仅 E4B + 短回复场景可加速（约 +37%）；E2B 或长回复反而变慢，默认关闭。
    • 底层修通了 Google 官方 iOS Metal sampler dylib 的隐藏符号问题：手写 Mach-O symbol walker 让引擎走 sampler_handles_input 优化路径而不是 fallback 慢路径。
    • 完整实测矩阵：E2B 中文 -27%、E2B 英文 -43%、E4B 英文短回复 +37%、E4B 中文长回复 -25%（内存压力）。
  • 🛠️ 下载器加诊断日志：mismatch 触发时会打印源标签、实际/期望字节数、容差范围，未来同类问题一眼能定位。

⚠️ 已知限制

  • LIVE 实时语音模式暂不可用：由于 Sideloadly 签名的 App 内存上限限制，LIVE 模式启动后会被系统终止。其他所有功能（Chat 对话、图片识别、语音识别、Agent 工具调用）均正常工作。
  • Sideloadly 签名 IPA 使用建议：受签名软件内存限制，E4B 模型只能用 CPU 推理（选 GPU 会报错），建议直接使用 E2B 模型，功能完备且更稳定。
  • MTP 推测解码注意：在 iPhone 上不是稳赢。E4B + 短回复有 ~37% 加速；E2B 或长回复（>500 token）会因 KV cache 撑大 + drafter 额外内存占用导致更慢。默认关闭，按需开启。
  • App 启动后需在设置中下载模型（约 2.4GB ~ 3.4GB），首次加载需等待约 30 秒。

📦 安装教程（Sideloadly 签名安装）

准备工作

需要 说明
iPhone iOS 17+，推荐 iPhone 15 Pro 及以上
Mac 或 Windows 电脑 用于签名
Apple ID 免费即可，不需要开发者账号
Sideloadly 下载安装到电脑
USB 数据线 连接手机和电脑

第一步：下载 IPA

从本页面的 Assets 区域下载 PhoneClaw-v1.3.2-unsigned.ipa 文件。

第二步：安装 Sideloadly

  1. 从 sideloadly.io 下载
  2. macOS：拖入「应用程序」
  3. Windows：运行安装程序

第三步：签名并安装到 iPhone

  1. 用 USB 连接 iPhone，在手机上点「信任此电脑」
  2. 打开 Sideloadly
  3. 拖动 PhoneClaw-v1.3.2-unsigned.ipa 到 Sideloadly 窗口
  4. 在左边面板选择你的 iPhone
  5. 输入你的 Apple ID（首次使用需要密码 + 双因素验证码）
  6. 点击 Start，开始签名安装
  7. 等待进度条走完，PhoneClaw 会出现在主屏幕

第四步：信任开发者证书

首次启动会提示「不受信任的开发者」，需要手动信任：

  1. 打开 iPhone 设置
  2. 进入 通用 → VPN 与设备管理（或「描述文件与设备管理」）
  3. 找到你的 Apple ID 邮箱所对应的开发者描述文件
  4. 点击 → 信任 "你的 Apple ID"
  5. 弹窗确认 → 点击 信任

第五步：启动 & 下载模型

  1. 回到主屏幕，打开 PhoneClaw
  2. 首次启动后，点击右上角 ⚙️ 配置 按钮
  3. 选择模型（推荐 Gemma 4 E2B，约 2.4GB — Sideloadly 签名下 E4B 在 GPU 会报错，只能 CPU）
  4. 点击 下载，等待下载完毕
  5. 返回对话页面，开始对话！

🔄 免费 Apple ID 限制

  • 免费 Apple ID 签名的 App 有效期 7 天，过期后需重新用 Sideloadly 安装
  • 每个免费 Apple ID 最多同时签 3 个 App
  • 模型文件保存在 App 沙盒内，重装后需重新下载

💡 功能亮点

  • 🗣️ 语音输入：录音后自动转为音频附件，模型直接分析语音内容
  • 📷 图片识别：拍照或选取图片，模型进行视觉理解
  • 🛠️ Agent 工具调用：联系人、剪贴板、日历等本地系统 Skills
  • 🔒 完全离线优先，本地处理：本地路径在设备端完成推理；聊天、图片和个人数据留在 iPhone；联网搜索、打开 URL 和 Mac 远程推理由用户显式触发

📱 设备兼容性

IPA 为 arm64 架构，理论上支持所有 iOS 17+ 的 iPhone。但模型加载需要足够内存：

设备 RAM 运行情况
iPhone 15 Pro / Pro Max 8GB ✅ 流畅
iPhone 16 / 17 全系列 8-12GB ✅ 流畅
iPhone 15 / 14 6GB ⚠️ 勉强，长对话可能被系统终止
iPhone XS ~ 13 4-6GB ❌ 内存不足，无法加载模型

推荐 iPhone 15 Pro 及以上机型以获得最佳体验。

PhoneClaw v1.3.2 — local AI agent for phones (English)

🧠 Gemma 4 LLM runs on-device by default. Text / image / voice multimodal chat; network use is limited to explicit user-triggered features.

🆕 What's New (v1.3.2)

  • 🐛 Fixed "download completes then auto re-downloads" loop: the previous downloader compared the just-finished file size against a hardcoded expectedFileSize constant, and a few-MB drift between the constant and the actual HuggingFace file size caused the validator to delete the just-completed download and trigger a redownload. The new validator only trusts HTTP-server-returned Content-Length / Content-Range. The paused-resume code path is patched for the same root cause.
  • ⚡ MTP speculative decoding toggle (experimental): a new switch in Settings to enable Gemma 4 Multi-Token Prediction speculative decoding. Only E4B + short replies see speedup (~+37%); E2B or long replies actually get slower. Off by default.
    • Under the hood we patched Google's iOS Metal sampler dylib's hidden-symbol issue: a hand-written Mach-O symbol walker makes the engine take the optimized sampler_handles_input path rather than the slow fallback.
    • Full empirical matrix: E2B Chinese -27%, E2B English -43%, E4B English short +37%, E4B Chinese long -25% (memory pressure under long outputs).
  • 🛠️ Downloader diagnostic logging: when the size validator fires, it now prints source label, actual vs expected bytes, and tolerance band — easier to triage similar issues in the future.

⚠️ Known Limitations

  • LIVE real-time voice mode is currently unavailable: Due to the memory cap on Sideloadly-signed apps, LIVE mode gets terminated by iOS on launch. All other features (chat, image recognition, voice recognition, agent tool calls) work normally.
  • Sideloadly-signed IPA usage note: Due to the sideload memory cap, the E4B model only works on CPU (GPU will fail). We recommend the E2B model — fully featured and more stable under the cap.
  • MTP speculative decoding caveats: not a guaranteed win on iPhone. E4B + short replies ≈ +37% speedup; E2B or long replies (>500 tokens) actually slow down because of KV-cache memory pressure plus the drafter's extra footprint. Off by default — enable per your scenario.
  • App requires a one-time model download (~2.4GB ~ 3.4GB) from Settings on first launch; initial load takes ~30 seconds.

📦 Installation (Sideloadly)

Requirements

Needed Details
iPhone iOS 17+, iPhone 15 Pro or newer recommended
Mac or Windows PC For signing
Apple ID Free account works; no developer account needed
Sideloadly Install on your computer
USB cable To connect iPhone to computer

Step 1: Download the IPA

Grab PhoneClaw-v1.3.2-unsigned.ipa from the Assets section below.

Step 2: Install Sideloadly

  1. Download from sideloadly.io
  2. macOS: drag into Applications
  3. Windows: run the installer

Step 3: Sign & install to your iPhone

  1. Connect your iPhone via USB, tap "Trust This Computer" on the phone
  2. Open Sideloadly
  3. Drag PhoneClaw-v1.3.2-unsigned.ipa into the Sideloadly window
  4. Select your iPhone in the left panel
  5. Enter your Apple ID (password + 2FA code required on first use)
  6. Click Start to sign and install
  7. Wait for the progress bar; PhoneClaw will appear on your home screen

Step 4: Trust the developer certificate

On first launch, iOS will prompt "Untrusted Developer". You need to trust it manually:

  1. Open Settings on iPhone
  2. Go to General → VPN & Device Management (or "Profiles & Device Management")
  3. Find the developer profile tied to your Apple ID email
  4. Tap it → Trust "Your Apple ID"
  5. Confirm the prompt → tap Trust

Step 5: Launch & download the model

  1. Open PhoneClaw from the home screen
  2. Tap the ⚙️ settings icon in the top-right
  3. Pick a model (Gemma 4 E2B recommended, ~2.4GB — E4B on GPU fails under Sideloadly, CPU only)
  4. Tap Download and wait for it to finish
  5. Return to the chat and start talking!

🔄 Free Apple ID Limits

  • Apps signed with a free Apple ID expire after 7 days — resign with Sideloadly to extend
  • A free Apple ID can sign up to 3 apps concurrently
  • Model files live in the app sandbox — reinstall requires re-download

💡 Feature Highlights

  • 🗣️ Voice input: Hold to talk, ASR transcribes and auto-sends
  • 📷 Image understanding: Take or pick a photo and ask questions
  • 🛠️ Agent tool calls: Contacts, clipboard, calendar, reminders, health, and more
  • 🔒 Fully offline first, private by default: The local path runs inference on-device; chat, images, and personal data stay on iPhone; Web Search, opening URLs, and Mac remote inference are explicit user-triggered capabilities

📱 Device Compatibility

The IPA is arm64. Supports all iOS 17+ iPhones in theory, but model loading needs RAM:

Device RAM Status
iPhone 15 Pro / Pro Max 8GB ✅ Smooth
iPhone 16 / 17 lineup 8-12GB ✅ Smooth
iPhone 15 / 14 6GB ⚠️ Tight — long chats may get terminated
iPhone XS ~ 13 4-6GB ❌ Insufficient memory

iPhone 15 Pro or newer recommended for the best experience.

Assets 3
Loading