# Luce-Org/lucebox

LLM speculative inference server for heterogeneous hardware & consumer GPUs

Repository: https://github.com/Luce-Org/lucebox
Canonical: https://ross.abutalabs.com/products/lucebox
Homepage: https://www.lucebox.com
Language: C++
License: Apache-2.0
License Family: permissive
Topics: kernel, llama-cpp, local-ai, qwen, rtx3090, megakernel, cuda, cuda-kernels, dflash, luce, speculative-decoding, pflash, speculative-prefill, poolside, spark, rocm, strix-halo, heterogeneous-computing, r9700
Last push: 2026-08-26T20:33:36+00:00

## Health v2 (maintenance only)
Score: 59/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 35, longevity 10
- inputs: {"age_days": 152, "days_push": 7, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, young
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2788, forks 265 (observed 2026-08-28T04:07:21.998804+00:00)

## What it is
Lucebox is an open-source LLM inference engine written in C++ with custom CUDA/HIP kernels, focused on speculative decoding and speculative prefill for heterogeneous hardware and consumer GPUs. It ships as a server with OpenAI- and Anthropic-compatible APIs and is tuned per model and hardware target (e.g., RTX 3090, Strix Halo, AMD R9700).

## Use cases
- run large language models locally on consumer GPUs
- speed up long-context prefill on a single GPU
- serve LLMs with an OpenAI-compatible API on self-hosted hardware
- combine unified memory and dedicated GPU for large-model inference
- get faster token generation with speculative decoding
- run inference across mixed CPU/GPU heterogeneous machines
- keep AI inference fully private on owned hardware

## When to choose
- you want maximum tokens/sec on consumer GPUs like RTX 3090 or AMD R9700
- you need fast time-to-first-token on very long prompts (64K-256K tokens)
- you want to mix unified-memory APUs with discrete GPUs in one inference setup
- you prefer a C++/CUDA engine with no Python, PyTorch, or Triton in the inference loop
- you want a plug-and-play private local AI server with compatible APIs

## When to avoid
- you need broad multi-vendor model support beyond the tuned model list
- you require multi-node distributed inference at datacenter scale
- you need a Python ecosystem for custom model experimentation
- you only need CPU-only inference without GPU acceleration
- you want a mature, widely-deployed engine like vLLM with extensive community tooling

## Facets
- artifact type: service
- maturity: active
- function: llm-inference, gpu-computing, http-server, sdk
- domain: large-language-models, machine-learning, gpu-computing, self-hosted, developer-tools
- platform: self-hosted, cpp
- tags: speculative-decoding, speculative-prefill, cuda-kernels, rocm, heterogeneous-computing, consumer-gpus, local-ai, openai-compatible-api, llama-cpp, moe-inference, paged-attention, long-context, linux, gpu, cuda

## Member repositories
- Luce-Org/lucebox (main) score 59

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:21.998804+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 2026-08-30T08:15:30.915873+00:00, confidence not recorded.
  - readme: https://github.com/Luce-Org/lucebox (fetched 2026-08-28T04:07:21.998804+00:00, sha 0aa143a8e864)
  - homepage: https://www.lucebox.com (fetched 2026-08-29T09:55:09.523806+00:00, sha 23754db01391)
  - site_page: https://www.lucebox.com/blog/pflash (fetched 2026-08-29T09:55:09.529649+00:00, sha ff894ee9f923)
- Data as of 2026-08-30T08:39:29.467469+00:00.