Magnitude — open-source self-tuning inference engine for local agents
让本机 Agent 推理引擎按任务调优。 Tune a local Agent inference engine for your tasks.
Launch HN https://github.com/magnitudedev/magnitude — Apache-2.0：面向本机 Agent 的推理引擎，在目标硬件上编译并调参 kernel，官网称相对 llama.cpp 最高约 2×（示例：M4 Pro Metal decode 30→57 tok/s；DGX Spark CUDA decode 49→58 tok/s，Qwen 3.6 35B A3B 4-bit / 64k / 无投机解码）。桌面端可一键接 Pi / OpenCode / Hermes / Codex / Claude Code 等，也提供 OpenAI 兼容接口；按需起停模型，会话间共享前缀缓存。晨间 changelog 把它与 Ledge 等一并标为「偏应用」跳过；晚间按「具体有用的 AI 工具」进 Cases 工具笔记。本条不替你下载模型或代跑推理。 The site claims up to about 2× vs llama.cpp (example: M4 Pro Metal decode 30→57 tok/s; DGX Spark CUDA decode 49→58 tok/s, Qwen 3.6 35B A3B 4-bit / 64k / no speculative decoding). Desktop app one-clicks Pi / OpenCode / Hermes / Codex / Claude Code and more, plus an OpenAI-compatible API; models start on demand and stop after idle; sessions share prefix caches. Morning changelog skipped it as app-shaped; this evening wave files it as a concrete AI-tool tip in Cases. This tip does not download models or run inference for you.
Magnitude 是 Apache-2.0 开源的本机推理引擎（仓库 magnitudedev/magnitude，站点 magnitude.dev）。面向「已经在跑 coding / 桌面 Agent、想把模型留在本机」的人：在你的 Mac / Linux / Windows 硬件上编译并调参 kernel，再按需起停模型。官网写可一键接 Pi、OpenCode、Hermes、Codex、Claude Code 等，也提供 OpenAI 兼容接口；会话之间共享前缀缓存，闲置后关掉模型。Launch HN item 49911995。本站收成 Cases「工具笔记」：补晨间跳过的具体有用工具，不是 Grok Bot 本体，也不是又一个通用聊天客户端。 Magnitude is an Apache-2.0 local inference engine (repo magnitudedev/magnitude, site magnitude.dev). It targets people already running coding / desktop agents who want models to stay on-device: compile and tune kernels on your Mac / Linux / Windows hardware, then start and stop models on demand. The site says one-click connects Pi, OpenCode, Hermes, Codex, Claude Code, and more, plus an OpenAI-compatible API; sessions share prefix caches and idle models shut down. Launch HN item 49911995. Listed here as a Cases tool tip filling a concrete gap the morning wave skipped — not Grok Bot itself, and not another generic chat client.
为什么推荐Why we recommend it 收录理由：本机跑 Agent 时推理引擎是真实痛点，且与 Pi/Codex/Claude Code 等已常见栈可对接。边界：官网/Launch HN 基准是厂商自测口径（HN 评论对 M5+/MLX 对比有异议）；本库未复跑；多 GPU / 部分机型检测仍有 issue。 Listing rationale: local-agent inference engines are a real pain point, and it plugs into stacks people already run (Pi/Codex/Claude Code). Boundary: site/Launch HN benches are vendor self-tests (HN comments dispute M5+/MLX comparisons); not re-run here; multi-GPU / some hardware detection still has open issues.
主源是 Launch HN 49911995、Apache-2.0 仓库与 magnitude.dev。预览 Playwright 实拍 magnitude.dev 首屏 3:2。不打星。 （item 49911995，2026-09-30T17:37:40Z / 约 2026-10-01 01:37 CST，Algolia 晚间核对 169pts） 仓库 magnitudedev/magnitude 于 2026-10-01 gh api：Apache-2.0、5926★； 站点 https://magnitude.dev/ HTTP 200。Primary sources are Launch HN 49911995, the Apache-2.0 repo, and magnitude.dev. Preview is a Playwright 3:2 capture of magnitude.dev above the fold. Unrated. (2026-09-30T17:37:40Z / about 2026-10-01 01:37 CST, 169pts on Algolia at the evening check) Launch HN item 49911995 at https://github.com/magnitudedev/magnitude — Apache-2.0 inference engine for local agents that compiles and tunes kernels on your hardware. Repo magnitudedev/magnitude on the 2026-10-01 gh api: Apache-2.0, 5926★; https://magnitude.dev/ HTTP 200.
左列是官网 / Launch HN / 仓库对得上的事实。右列说明为什么放在 Cases 工具笔记，以及不要把它当成已经替你验证过的「全面碾压 llama.cpp / MLX」。 Left column: what the site / Launch HN / repo actually say. Right column: why it sits as a Cases tool tip, and why it is not a verified “beats llama.cpp / MLX everywhere” claim.
Magnitude Magnitude
本机推理引擎 + 桌面端：为目标硬件调 kernel，给 Agent 提供本地 OpenAI 兼容端点。 Local inference engine + desktop app: tunes kernels to your hardware and serves an OpenAI-compatible endpoint for agents.
对照本站口径 This library’s framing
Cases「AI tool tip」。帮本机 Agent 跑开源权重，不是云端 Bot 托管。 An AI tool tip in Cases — helps local agents run open weights; not hosted cloud Bot compute.
Magnitude Magnitude
官网示例相对 llama.cpp：Metal decode +92%，CUDA decode +19%（给定模型与设置）。 Site examples vs llama.cpp: Metal decode +92%, CUDA decode +19% (given model and settings).
对照本站口径 This library’s framing
HN 评论对 M5+ / MLX 对比有异议；本库未复跑，选型以你机器实测为准。 HN comments dispute M5+ / MLX comparisons; not re-run here — measure on your machine before choosing.
Magnitude Magnitude
官网列 Pi / OpenCode / Hermes / Codex / Claude Code 等一键连接；也可走 OpenAI 兼容 API。 Site lists one-click Pi / OpenCode / Hermes / Codex / Claude Code and more; OpenAI-compatible API also works.
对照本站口径 This library’s framing
不替代 Cursor Cloud Agent / Grok Bot 云电脑；本条也不代你改 Agent 配置。 Does not replace Cursor Cloud Agents / Grok Bot cloud computers; this tip also does not rewrite your agent config for you.
打开 https://magnitude.dev/ 与 HN item 49911995，核对 Apache-2.0、支持的 OS / Agent、以及他们自己贴的基准表。再决定是否值得在你硬件上试。 Open https://magnitude.dev/ and HN item 49911995; confirm Apache-2.0, supported OS / agents, and their own bench table before deciding a trial on your hardware.
从官网或仓库 README 下载对应平台安装包。本站不代装。装好后先看模型目录与硬件检测结果，再决定拉哪一档量化。 Download the platform installer from the site or repo README. This site does not install it. After install, read the model catalog and hardware detection before picking a quant.
用官网列出的一键连接，或把 OpenAI 兼容基址指到 Magnitude。用你真实的长上下文 / 工具调用轨迹测 decode 与显存，不要只抄宣传数字。密钥与本地仓库权限仍按你自己的审批边界。 Use a listed one-click connector, or point an OpenAI-compatible base URL at Magnitude. Measure decode and memory on your real long-context / tool-call traces — do not copy marketing numbers. Keep your own approval boundary for secrets and repo write access.
