# xLLM-AI/xllm

A high-performance inference engine for LLM, VLM, DiT and REC models, optimized for diverse AI accelerators. It is hosted in OpenAtom Foundation.

Repository: https://github.com/xLLM-AI/xllm
Canonical: https://ross.abutalabs.com/products/xllm
Homepage: https://xllm-ai.com/
Language: C++
License: Apache-2.0
License Family: permissive
Topics: deepseek, inference, inference-engine, llm-inference, qwen, large-language-models, glm
Last push: 2026-08-24T09:51:10+00:00

## Health v2 (maintenance only)
Score: 78/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 99, release rhythm 81, longevity 27
- inputs: {"age_days": 386, "days_push": 9, "days_rel": 50, "gap_med": 34, "n_releases_24m": 10}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1536, forks 285 (observed 2026-08-28T04:05:00.069176+00:00)

## What it is
xLLM is a high-performance C++ inference engine for LLM, VLM, DiT and recommendation models, optimized for heterogeneous AI accelerators such as Ascend NPUs, Cambricon MLUs, Iluvatar CoreX and Moore Threads GPUs. It provides a unified service-engine stack with asynchronous scheduling, graph optimization, and global KV cache management, and is hosted under the OpenAtom Foundation.

## Use cases
- serve deepseek or qwen models on ascend npu hardware
- run high-throughput llm inference on domestic chinese accelerators
- deploy vision-language models like glm-4.6v for inference
- reduce inference cost with efficient kv cache management
- serve moe models like deepseek-v4 with day-0 support
- self-host an openai-compatible llm serving endpoint

## When to choose
- you need to run large language models on Chinese domestic accelerators like Ascend or Cambricon
- you want day-0 support for new open models such as DeepSeek, Qwen, or GLM releases
- you need high-throughput, low-latency distributed inference with MoE optimizations
- you want a unified engine for LLM, VLM, and DiT model serving

## When to avoid
- your hardware is limited to NVIDIA GPUs where vLLM or SGLang have broader ecosystem support
- you need a pure training framework rather than an inference engine
- you require extensive community plugins and third-party integrations available in more mature engines

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, gpu-computing, machine-learning, http-server
- domain: large-language-models, deep-learning, artificial-intelligence, gpu-computing
- platform: cpp, self-hosted
- tags: inference-engine, npu, ascend, cambricon, vlm, moe, kv-cache, serving, linux, docker, gpu

## Member repositories
- xLLM-AI/xllm (main) score 78

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:00.069176+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 2026-08-30T04:31:07.051640+00:00, confidence not recorded.
  - readme: https://github.com/xLLM-AI/xllm (fetched 2026-08-28T04:05:00.069176+00:00, sha 8cec202aac18)
  - homepage: https://xllm-ai.com/ (fetched 2026-08-29T11:33:00.055160+00:00, sha 2be4eabc0ce8)
  - site_page: https://docs.xllm-ai.com/en (fetched 2026-08-29T11:33:00.064108+00:00, sha bca54149aeca)
- Data as of 2026-08-30T08:39:29.467469+00:00.