← Runners

MLX（mlx-lm）MLX（mlx-lm）

Apple

开源Open source MITMIT Apple硅Apple silicon 统一内存Unified memory mlx-lmmlx-lm partial lifecycle: active 核验Checked 2026-10-30

MLX 是 Apple 硅的官方本地推理栈：统一内存架构的原生框架（mlx）+ 官方 LLM 运行线（mlx-lm）， MIT、活跃维护。Mac 用户跑本地模型的「亲儿子」路线——连微调都在同一栈里。

MLX is Apple's official local-inference stack for Apple silicon: a native framework for unified memory (mlx) plus the official LLM layer (mlx-lm), MIT, actively maintained. The 'native son' route for Mac users — even fine-tuning lives in the same stack.

适合与不适合When to use it, when not to

适合：M 系 Mac 要原生性能与统一内存红利、要在 Mac 上做 LoRA 微调、mlx-community 模型生态。不适合：非 Apple 硬件（框架专属）；要跨平台单一技术栈。

Fits: native performance on M-series Macs (unified-memory dividend), LoRA fine-tuning on Mac, the mlx-community model ecosystem. Not a fit: non-Apple hardware; one cross-platform stack.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-21

出品属地Origin: 国外出品foreign-made　Apple（美国）。Apple (US).
国内使用Direct access: 大陆可直连direct OK　ml-explore.github.io 直连实测 200（2026-10-21）。ml-explore.github.io measured 200 direct (2026-10-21).
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-21 检索）；开源项目未面向大陆公众提供服务，备案不适用。not in the CAC public registry (searched 2026-10-21); the open-source project does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: MLX 原生格式 + HF 转换（mlx-community 生态）；GGUF 可转换（mlx-community 有转换线）；模型清单随社区滚动，未逐条清点。
MLX-native format + HF conversion (the mlx-community ecosystem); GGUF convertible (mlx-community has a conversion line); the model list rolls with the community, not inventoried item by item.
运行位置Where it runs: Apple 硅（M 系）统一内存架构的原生张量框架；mlx-lm 是官方 LLM 运行线（「Run LLMs with MLX」仓库描述直抓所见）。GitHub mlx 28,975★ / mlx-lm 7,506★（API 核验 2026-10-21）。
A native tensor framework for the Apple-silicon (M-series) unified-memory architecture; mlx-lm is the official LLM runtime line ("Run LLMs with MLX" repo description, direct-fetched). GitHub mlx 28,975★ / mlx-lm 7,506★ (API verified 2026-10-21).
本地文件Local files: 全本地：模型与输出都在自有环境；未见官方云线——Apple 未提供配套云端推理线（未见声明）。
Fully local: models and outputs live in your own environment; no official cloud line seen — Apple provides no companion cloud-inference line (no statement seen).
关机后的任务Tasks after shutdown: mlx_lm.server 提供 OpenAI 兼容服务线；细节未逐条核验。
mlx_lm.server provides an OpenAI-compatible service line; details not verified item by item.
工具与扩展Tools & extensions: mlx-lm 含 tool calling 线；MCP 支持未逐条核验。
mlx-lm includes a tool-calling line; MCP not verified item by item.
上下文与记忆Context & memory: 上下文按参数配置；KV 缓存量化在 mlx-lm 文档线，细节未逐条核验。
Context is set by parameters; KV-cache quantization is in the mlx-lm docs line, details not verified item by item.
权限与限制Permissions & limits: 本地进程无账号体系；命令行/服务线鉴权由使用者自配（mlx_lm.server 细节未逐条核验）。
A local process with no account system; CLI/service-line authentication configured by the user (mlx_lm.server details not verified item by item).
适合什么任务Best for: 适合：Apple 硅机器要原生性能（统一内存红利）、要在 Mac 上微调（mlx-lm 带微调线）。不适合：非 Apple 硬件；要跨平台同一栈（llama.cpp 覆盖面更宽）。
Fits: Apple-silicon machines wanting native performance (the unified-memory dividend), fine-tuning on the Mac (mlx-lm carries a fine-tuning line). Does not fit: non-Apple hardware; wanting one cross-platform stack (llama.cpp covers more ground).

头号误解Top misconception

把 MLX 当通用推理引擎——Apple 硅专属，x86/NVIDIA 不是它的场Treating MLX as a general inference engine — it is Apple-silicon exclusive; x86/NVIDIA is not its arena
只记 mlx 框架——跑 LLM 要用 mlx-lm 这层Remembering only the mlx framework — running LLMs needs the mlx-lm layer
把 mlx-community 模型当官方出品——社区转换生态，质量随贡献者Treating mlx-community models as official — a community conversion ecosystem whose quality varies by contributor

未知项清单Known unknowns

mlx-community 模型全清单与转换质量（社区生态，未逐条清点）
mlx_lm.server 的 OpenAI 兼容面细节（未逐条核验）
MCP 支持状态（未核验到官方声明）

The full mlx-community model list and conversion quality (a community ecosystem, not inventoried item by item)
Details of mlx_lm.server's OpenAI-compatibility surface (not verified item by item)
MCP support status (no official statement verified)

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-30，内容更新日content updated on 2026-10-21。

类型Type	名称Name	链接Link
repo	ml-explore/mlx · 仓库（MIT；28,975★，最近推送 2026-10-30，GitHub API 核验）ml-explore/mlx · repo (MIT; 28,975★, last push 2026-10-30, GitHub API verified)	https://github.com/ml-explore/mlx
repo	ml-explore/mlx-lm · 仓库（MIT；7,506★，最近推送 2026-10-29，GitHub API 核验）ml-explore/mlx-lm · repo (MIT; 7,506★, last push 2026-10-29, GitHub API verified)	https://github.com/ml-explore/mlx-lm
docs	MLX · 官方文档MLX · official docs	https://ml-explore.github.io/mlx/build/html/index.html
changelog	ml-explore/mlx · Releases（changelog）ml-explore/mlx · Releases (changelog)	https://github.com/ml-explore/mlx/releases

实测记录Measured runs

本档未跑吞吐与显存实测；结论只覆盖运行时能力与许可。

This site has not completed measured runs. The closest test protocol is the code-side protocols/framework/_protocol.md (the local-endpoint and credential-boundary sections are reusable; no runner-specific protocol exists on this site yet).