# GradientHQ/parallax

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

Repository: https://github.com/GradientHQ/parallax
Canonical: https://ross.abutalabs.com/products/gradienthq-parallax
Language: Python
License: Apache-2.0
License Family: permissive
Topics: chatbot, distributed-systems, kimi, llama, llm, qwen, oss-gpt, large-language-models, python, pytorch, decentralized-inference, blackwell, transformer, llm-serving, glm, minimax, deepseek
Last push: 2026-07-01T07:35:43+00:00

## Health v2 (maintenance only)
Score: 66/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 90, release rhythm 59, longevity 24
- inputs: {"age_days": 345, "days_push": 63, "days_rel": 274, "gap_med": 14, "n_releases_24m": 4}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1368, forks 146 (observed 2026-08-28T04:04:31.562781+00:00)

## What it is
Parallax is a distributed model serving framework that lets you build your own AI cluster across heterogeneous, geographically dispersed devices. It shards LLMs via pipeline parallelism and uses P2P communication to serve models like Llama, Qwen, and DeepSeek on personal hardware.

## Use cases
- host a local LLM across my own devices
- build a private AI cluster from spare machines
- serve large models that don't fit on one GPU
- run decentralized LLM inference without a datacenter
- shard an LLM across Mac and Linux nodes
- self-host a chatbot backend on consumer hardware

## When to choose
- you want to pool GPUs across multiple machines or locations to serve large LLMs
- you need decentralized, self-hosted inference without centralized cloud infrastructure
- your nodes have heterogeneous hardware and you need dynamic scheduling and routing

## When to avoid
- you need a single-GPU, single-node inference server with minimal setup
- you require battle-tested, production-grade serving at massive scale
- you need low-latency inference where P2P network hops would hurt performance

## Facets
- artifact type: framework
- maturity: experimental
- function: llm-inference, agent-framework, chatbot, rpc, scheduling
- domain: large-language-models, microservices, artificial-intelligence, self-hosted, machine-learning
- platform: python, cross-platform, self-hosted
- tags: decentralized-inference, pipeline-parallelism, model-sharding, p2p-networking, llm-serving, ai-cluster, kv-cache, continuous-batching, macos, linux, docker

## Member repositories
- GradientHQ/parallax (main) score 66

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:31.562781+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:41:01.961919+00:00, confidence not recorded.
  - readme: https://github.com/GradientHQ/parallax (fetched 2026-08-28T04:04:31.562781+00:00, sha 512c9c123c05)
- Data as of 2026-08-30T08:39:29.467469+00:00.