AI basics

Start here for Mind. No serving cluster required.

Check the host

buraaq --version
buraaq ai doctor

Doctor reports CPU, RAM, GPUs, CUDA hints, backend availability (llama.cpp, vLLM, OpenAI-compatible), cache path, Strata cache/free disk, and whether the train env is ready.

Inspect a model

buraaq ai inspect Qwen/Qwen3-8B
buraaq ai strata Qwen/Qwen3-32B

You get parameter estimate, context, and VRAM estimates (FP16 / INT8 / 4-bit). Numbers are planner estimates until the model is cached with a real config.json. For models larger than VRAM, see Strata.

buraaq ai pull Qwen/Qwen3-8B

Pull stores metadata under ~/.buraaq/models (and full weights if huggingface-cli / hf is on PATH).

One-shot and chat

Needs a reachable backend: llama-server, vLLM, or BURAAQ_AI_BASE_URL pointing at an OpenAI-compatible server.

buraaq ai run Qwen/Qwen3-8B -p "Explain ownership in Buraaq"
buraaq ai chat Qwen/Qwen3-8B

The planner prints why it chose quant and backend before starting.

From Buraaq source

fn main() {
    m = ai.model("Qwen/Qwen3-8B")
    ai.system(m, "You are concise.")
    println(ai.chat(m, "Say hello in one sentence."))
}

std.ai talks HTTP to BURAAQ_AI_BASE_URL (default http://127.0.0.1:8000). Start serve first (next page), or point at any compatible upstream.

Commands at this level

Next: Serve models or Strata.