# AIMA

> AIMA (灵机) — "AI Infrastructure, managed by AI." A single Go binary that runs LLM inference on your hardware. AIMA detects the accelerator (NVIDIA / AMD / Huawei Ascend / Hygon DCU / Moore Threads / MetaX / Apple Silicon, or CPU-only), picks an inference engine (vLLM / SGLang / llama.cpp) and configuration from a YAML knowledge base, deploys the model, runs benchmarks, and writes the winning config back. The loop is driven by a built-in PDCA agent (codename Explorer). AIMA is also an MCP server — any MCP-compatible runtime can drive it directly. It runs in production as the inference backend behind OpenClaw, covering LLM, ASR, TTS, image generation, and VLM. Goal: Ollama-level TCO with SOTA inference performance. Open source under Apache 2.0.
>
> There is also AIMA Cloud (灵机云) — an invite-only beta cloud service where a cloud-side AI agent connects to your devices and installs, diagnoses, repairs, and upgrades them remotely, including one-command install of Dify, ComfyUI, Open WebUI, and OpenClaw.

## How it works

1. Install (macOS / Linux): `curl -fsSL https://raw.githubusercontent.com/Approaching-AI/AIMA/master/install.sh | sh`
2. Install (Windows PowerShell): `irm https://raw.githubusercontent.com/Approaching-AI/AIMA/master/install.ps1 | iex`
3. Or download a pre-built binary from the Releases page, or build from source.
4. `aima hal detect` — see the detected accelerator, driver versions, and RAM.
5. `aima onboarding` — read-only first-run guide; recommends the next command.
6. `aima run qwen3-4b` — resolves the model, picks engine + config for this host, pulls assets, deploys.
7. `aima serve` — OpenAI-compatible API + Web UI at http://localhost:6188.
8. Linux shared server path: `sudo aima init` → `aima deploy <model>` → `aima serve`.

## Engines

- vLLM — auto-selected for discrete NVIDIA / AMD / Ascend / DCU / Moore Threads / MetaX GPUs
- SGLang — auto-selected where throughput benchmarks favour it
- llama.cpp — auto-selected for Apple Silicon, CPU-only, and low-VRAM scenarios

Engine selection is automatic per hardware — no manual flags required.

## Validated silicon

- NVIDIA (CUDA)
- AMD (ROCm)
- Huawei Ascend (CANN)
- Hygon DCU
- Moore Threads (MUSA)
- MetaX (MACA)
- Apple Silicon (Metal / CPU)
- Intel CPU-only

Every release tag goes through a UAT matrix on real hardware across all vendors listed above. Benchmark data for NVIDIA GB10 / DGX Spark: see /benchmarks.

## MCP

Point any MCP-compatible client at AIMA's HTTP endpoint:

```jsonc
{
  "mcpServers": {
    "aima": { "type": "http", "url": "http://<aima-host>:6188/mcp" }
  }
}
```

The MCP surface covers: hardware detection, model scan, engine selection, deployment, benchmark, LAN fleet discovery, knowledge sync.

## AIMA Cloud (灵机云)

Invite-only beta. A cloud-side AI agent connects to your device and can install, diagnose, repair, and upgrade it remotely.

Connect a device:

```bash
curl -sL https://aimaservice.ai/go | bash
# enter your invite code when prompted
```

Supported one-command installs via cloud agent: Dify, ComfyUI, Open WebUI, OpenClaw.

## Links

- Website (Global): https://aimaservice.ai
- Website (China): https://aimaserver.com
- AIMA (open source): https://github.com/Approaching-AI/AIMA
- AIMA Cloud device CLI (open source): https://github.com/Approaching-AI/aima-server
- Product page: /product
- Benchmarks: /benchmarks
- Cloud: /cloud
- Blog: /blog
- Releases: https://github.com/Approaching-AI/AIMA/releases
- Full LLM-readable details: /llms-full.txt