登录

parallax

GradientHQApache-2.0更新放缓

分布式模型服务框架，能把多台机器组成自建 AI 推理集群，用来跑 Llama、DeepSeek 等大模型。

Parallax is a distributed model serving framework that lets you build your own AI cluster anywhere

v0.1.2发布于 2025-12-02 07:53 UTC1,421star152fork同步于 2026-11-07 07:31 UTC
分布式推理模型服务推理集群大模型部署去中心化
对比 GitHub 官网blackwellchatbotdecentralized-inferencedeepseekdistributed-systems

怎么上手

由 README 提炼，执行前请确认
原文 
git clone https://github.com/GradientHQ/parallax.git
cd parallax
./install.sh
source .venv/bin/activate
parallax serve -m Qwen/Qwen3.5-0.8B
硬件要求按 README 说明

README 未说明具体显存与 GPU 型号要求；支持跨平台部署，GPU 走 SGLang/vLLM 后端，Mac 走 MLX LM 后端。

支持 Apple Silicon
下载：源码 zip·GitHub Releases（安装包）本站不镜像安装包

项目说明

由 README 自动翻译提炼
没人工点评？写下你的经验

分布式模型推理框架，把配置各异的设备组成你自己的 AI 集群

分布式模型服务框架，能把多台机器组成自建 AI 推理集群，用来跑 Llama、DeepSeek 等大模型。

核心能力

  • 用流水线并行把模型分片到多个节点
  • Mac 端用 Paged KV cache 与连续批处理
  • 动态请求调度与路由提升吞吐
  • P2P 通信由 Lattica 支撑跨节点互联
  • GPU 后端接 SGLang/vLLM，Mac 后端接 MLX LM

容易踩的坑

安装脚本会创建 .venv，需先 source 激活才能用 parallax 命令；首次运行需从 HuggingFace 拉取模型权重；GPU 与 Mac 后端分别依赖 SGLang/vLLM 和 MLX LM，需按平台准备对应环境。

内容由模型根据 README 生成，可能不完整或不准确，以官方文档为准。看原文 README

最近发版

4 个版本v0.1.2
全部 release
  1. v0.1.2最新正式版2025-12-02 · 2025-12-02 07:53 UTC

    Highlights

    • Add paged kv implement
    • Support switch model in chat page

    What's Changed

    Full Changelog: https://github.com/GradientHQ/parallax/compare/v0.1.1...v0.1.2

  2. v0.1.12025-11-26 · 2025-11-26 04:17 UTC
    查看更新说明

    Highlights

    • Add docker image for dgx spark
    • Support huggingface offline
    • Refactor executor and p2p server use sub-process instead of sub-thread
    • Add paged attn naive kernel

    What's Changed

  3. v0.1.02025-11-11 · 2025-11-11 12:29 UTC
    查看更新说明

    Highlights

    • Support for Kimi-K2, DeepSeek, MiniMax, GLM and gpt-oss-safeguard
    • Optimize communication methods and add relay server support for remote connections
    • Add parallax chat command to run a chat page server
    • Adjust the readme structure and improve the user guide and contributing guide

    What's Changed

  4. v0.0.12025-09-29 · 2025-09-29 10:03 UTC

    这个版本没有写更新说明。

用户点评

登录用户的真实使用记录，可按版本和硬件筛选
共 0 条

    还没有人点评。第一个说出真实体验的人最有价值。

    README 原文

    未翻译
    在 GitHub 打开

    license issue resolution open issues

    Gradient | Blog | X/Twitter (Gradient) | X/Twitter (Parallax) | Discord | Arxiv

    News

    • [2026/02] 🦞 Parallax now supports OpenClaw integration! See Docs
    • [2025/10] 🔥 Parallax won #1 Product of The Day on Product Hunt!
    • [2025/10] 🔥 Parallax version 0.0.1 has been released!

    About

    A fully decentralized inference engine developed by Gradient. Parallax lets you build your own AI cluster for model inference onto a set of distributed nodes despite their varying configuration and physical location. Its core features include:

    • Host local LLM on personal devices
    • Cross-platform support
    • Pipeline parallel model sharding
    • Paged KV cache management & continuous batching for Mac
    • Dynamic request scheduling and routing for high performance

    The backend architecture:

    User Guide

    Quick Install

    git clone https://github.com/GradientHQ/parallax.git
cd parallax
./install.sh
source .venv/bin/activate
parallax serve -m Qwen/Qwen3.5-0.8B

    Contributing

    We warmly welcome contributions of all kinds! For guidelines on how to get involved, please refer to our Contributing Guide.

    Supported Models

    ProviderHuggingFace CollectionBlogDescription
    DeepSeekDeepseekDeepSeek-V3.2DeepSeek-R1 DeepSeek-V3.2 ReleaseDeepSeek-V3.2 is built for stronger reasoning and agent workflows, adding thinking-aware tool use while keeping both thinking and non-thinking tool-use modes available.
    MiniMaxMiniMax AIMiniMax-M3MiniMax-M2.7MiniMax M3: Frontier Coding, 1M Context, Native MultimodalityMiniMax-M3 is a sparse-attention model for coding, agentic work, long-context tasks, and multimodal input, with MiniMax-M2.7 kept as a supported sibling model.
    GLMZ AIGLM-5.2GLM-5.1GLM-4.7GLM-5.2: Built for Long-Horizon TasksGLM-5.2 is Z AI's latest flagship model for long-horizon coding agents, large-scale implementation, automated research, and performance optimization.
    Kimi-K2Moonshot AIKimi-K2-ThinkingKimi-K2-Instruct-0905Kimi K2 ThinkingKimi-K2-Thinking is the reasoning-focused Kimi-K2 variant for long-horizon agentic workflows, multi-step reasoning, and tool-heavy tasks.
    QwenQwenQwen3.6-35B-A3BQwen3.6-35B-A3B: Agentic Coding Power, Now Open to AllQwen3.6-35B-A3B is an open-weight MoE model focused on stable, practical coding and agentic workflows, with Qwen3-Next and larger Qwen3 MoE variants also represented.
    gpt-ossOpenAIgpt-oss-120bgpt-oss-safeguard-120bIntroducing gpt-oss-safeguardgpt-oss-120b is OpenAI's open-weight GPT model; gpt-oss-safeguard-120b adds policy-driven reasoning classification for flexible moderation and safety workflows.
    StepStepFunStep-3.5-FlashStep 3.5 FlashStep-3.5-Flash is a sparse MoE foundation model for efficient reasoning and agentic workflows, activating a small subset of its total parameters per token.

    相关项目

    同分类里 star 最高的项目
    对比全部 →
    ollama
    ollama·Go

    本地一键跑起 Kimi、DeepSeek、Qwen、Gemma 等大模型，省去环境配置和依赖折腾，适合快速做推理和原型验证。

    本地推理+3大模型推理服务入门
    186510+32019kv0.40.152026-11-10 01:49 UTC
    llama.cpp
    ggml-org·C++

    用 C/C++ 实现的大模型推理引擎，能在普通 CPU 和消费级显卡上跑 LLaMA 等模型，支持量化以降低显存占用。

    推理引擎+4大模型推理服务入门
    133898+30125kb122332026-11-10 10:59 UTC
    vllm
    vllm-project·Python

    面向大语言模型的高吞吐、低显存推理与服务引擎，通过 PagedAttention 等优化显著提升并发吞吐并降低显存占用，适合部署在线推理服务。

    推理引擎+4大模型推理服务进阶
    96670+17725kv0.33.0rc32026-11-10 11:13 UTC
    jan
    janhq·Rust

    在本地电脑上离线运行大模型的桌面客户端，是 ChatGPT 的开源替代品，无需联网即可对话，数据不出本机。

    本地推理+4大模型推理服务
    45651+1093.2kv0.8.82026-11-10 09:10 UTC