vLLM

The serving engine production runs on. PagedAttention and continuous batching, on your own GPU.

vLLM is the serving engine production actually uses: PagedAttention, continuous batching, an OpenAI-compatible server on localhost:8000/v1. You bring the GPU.

`vllm serve Qwen/Qwen3-8B` and you are done. CUDA or ROCm; on a Mac reach for MLX instead. No quota, no key, nothing leaving the machine.

Max context: Set with --max-model-len
Free tier: Unlimited and private
Requirement: Nothing — no account

Endpoint

Base URLhttp://localhost:8000/v1

Rate limitNone — it is your machine

OpenAI-compatible
text
vision

Code examples

npm install openai

import OpenAI from 'openai'

const client = new OpenAI({
  baseURL: 'http://localhost:8000/v1',
  apiKey: 'not-needed'
})

const response = await client.chat.completions.create({
  model: 'Qwen/Qwen3-8B',
  messages: [{ role: 'user', content: 'Explain closures in one paragraph.' }]
})

console.log(response.choices[0].message.content)

pip install openai

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="not-needed",
)

response = client.chat.completions.create(
    model="Qwen/Qwen3-8B",
    messages=[{"role": "user", "content": "Explain closures in one paragraph."}],
)

print(response.choices[0].message.content)

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3-8B",
    "messages": [{"role": "user", "content": "Explain closures in one paragraph."}]
  }'

Worth knowing

Start the server first: `vllm serve Qwen/Qwen3-8B`.
Needs a CUDA or ROCm GPU; on a Mac reach for MLX instead.

Models you get for free5

Qwen 3qwen3
Llama 3.3 70Bllama-3.3-70b
GPT-OSS 120Bgpt-oss-120b
DeepSeek R1deepseek-ai/DeepSeek-R1
DeepSeek V4deepseek-v4-flash