Rapid-MLX 1.0:Apple Silicon 本地推理服务加入内置 Agent 挽具与新模型族
开源项目 Rapid-MLX 于 10 月 29 日发布 1.0.0:在 Apple Silicon 上提供兼容 OpenAI/Anthropic API 的本地推理,并加入终端与桌面内置 Agent 挽具,同时引入 LFM3、Ornith 与 MiniMax-H3 等模型别名。
面向编码智能体的本地推理栈 Rapid-MLX 在 2026 年 10 月 29 日放出稳定版 1.0.0。项目基于 Apple 的 MLX,在 Apple Silicon Mac 上提供 OpenAI 与 Anthropic 兼容 API,许可证为 Apache-2.0;官方与 PyPI 均强调其工具调用面向编码智能体场景。
1.0 的核心增量是内置 Agent 挽具:终端可用 rapid-mlx chat --harness 维持带工具与审批的会话,桌面端同能力默认关闭,需在设置中打开「Built-in Agent Harness」。网络客户端需审批,终端命令继承受限环境,但官方写明策略偏启发式、命令本身未沙箱化。模型侧增加 LFM3 1.7B 文本与 LFM3-VL 2B 图像的 8-bit 别名、Ornith 1.5 122B-A10B 的 4-bit 文本别名,以及可用文本或首帧生成带同步音频视频的 MiniMax-H3(官方称服务端需约 160GB 及以上统一内存与 BF16 权重,默认 8 步 Turbo LoRA)。桌面聊天单条最多附 8 张图,并可在空闲后释放模型以腾内存。
本地推理无需账号;若开启联网工具则会出站。分享算力需 QuickSilver 提供方密钥,回报为 API 积分而非现金。桌面要求 macOS 14+、Apple Silicon,16GB 内存为更舒适基线。对希望把 Claude Code、Codex 一类智能体指到本机模型、少把提示词送上云的开发者,1.0 把「先搭推理服务再接 Agent」收成更短的路径,但仍需按机型内存与模型页要求选型。
Rapid-MLX 1.0: local Apple Silicon inference with built-in agent harness
🧠 A coding agent pointed at a cloud model sends every prompt off your Mac. Running the model locally usually means assembling an inference server before the agent can start.
Meet Rapid-MLX from @rapidmlx: a free, Apache-2.0 inference server for Apple silicon Macs, built on MLX with OpenAI- and Anthropic-compatible APIs. Stable 1.0.0, released 29 October, adds a built-in agent harness and support for new model families.
🧰 Keep the agent loop on the Mac. rapid-mlx chat --harness gives you persistent agent chats with tools and approvals, image attachments and measured model-loading progress. The Desktop app offers the same chats behind the Built-in Agent Harness setting, which is off by default. Network clients require approval and terminal commands inherit a restricted environment, though the policy is heuristic and those commands are not sandboxed.
🎛️ Choose a model that fits the memory. Native LFM3 1.7B text and LFM3-VL 2B image support use pinned 8-bit aliases, while Ornith 1.5 122B-A10B gets a pinned 4-bit text alias. MiniMax-H3 generates video with synchronised audio from text or a first image; serving it needs 160 GB or more of unified memory and BF16 weights, and defaults to the 8-step Turbo LoRA.
🖼️ Put more context in one message. A Desktop chat now accepts up to 8 images, and small models are reminded of their tools during long conversations. An optional setting releases idle models after a chosen idle time, freeing memory for the next task.
The Desktop app needs macOS 14 or later on an Apple silicon Mac, with 16 GB of memory as a comfortable baseline. Turn on the Built-in Agent Harness in settings before pointing an agent at a chat, and check the model's memory requirement before downloading it.
Free under Apache-2.0. Sharing compute needs a QuickSilver provider key and earns API credit, not a cash payout.
Find it on OpenNavo: https://t.co/iHL3kfjX09
#MacApps #LocalAI #AppleSilicon #OpenSource