Build your own AI model inference engines

2 hours ago 2

The framework to build your own inference engine with expert control.
Build inference APIs, agents, MCP servers, RAG, and pipelines.
No MLOps. No YAML.

Over 340,000 developers use Lightning Cloud - purpose-built for LitServe, PyTorch and PyTorch Lightning. Get GPUs from $0.19, frontier-grade training/inference clusters, vibe training/inference workspaces, notebooks with AI assistance and inference for your custom models.

LitServe lets you build your own inference engine. Serving engines such as vLLM serve specific model types (LLMs) with rigid abstractions. LitServe gives you the low-level control to serve any model (vision, audio, text, multi-modal), and define exactly how inference works - from batching, caching, streaming, and routing, to multi-model orchestration and custom logic. LitServe is perfect for building inference APIs, agents, chatbots, MCP servers, RAG, pipelines and more.

Self host LitServe or deploy in one-click to Lightning AI.

Install LitServe via pip (more options):

Example 1: Toy inference pipeline with multiple models.
Example 2: Minimal agent to fetch the news (with OpenAI API).
(Advanced examples):

Inference pipeline example

import litserve as ls # define the api to include any number of models, dbs, etc... class InferencePipeline(ls.LitAPI): def setup(self, device): self.model1 = lambda x: x**2 self.model2 = lambda x: x**3 def predict(self, request): x = request["input"] # perform calculations using both models a = self.model1(x) b = self.model2(x) c = a + b return {"output": c} if __name__ == "__main__": # 12+ features like batching, streaming, etc... server = ls.LitServer(InferencePipeline(max_batch_size=1), accelerator="auto") server.run(port=8000)

Deploy for free to Lightning cloud (or self host anywhere):

# Deploy for free with autoscaling, monitoring, etc... lightning deploy server.py --cloud # Or run locally (self host anywhere) lightning deploy server.py # python server.py

Test the server: Simulate an http request (run this on any terminal):

curl -X POST http://127.0.0.1:8000/predict -H "Content-Type: application/json" -d '{"input": 4.0}'

import re, requests, openai import litserve as ls class NewsAgent(ls.LitAPI): def setup(self, device): self.openai_client = openai.OpenAI(api_key="OPENAI_API_KEY") def predict(self, request): website_url = request.get("website_url", "https://text.npr.org/") website_text = re.sub(r'<[^>]+>', ' ', requests.get(website_url).text) # ask the LLM to tell you about the news llm_response = self.openai_client.chat.completions.create( model="gpt-3.5-turbo", messages=[{"role": "user", "content": f"Based on this, what is the latest: {website_text}"}], ) output = llm_response.choices[0].message.content.strip() return {"output": output} if __name__ == "__main__": server = ls.LitServer(NewsAgent()) server.run(port=8000)

Test it:

curl -X POST http://127.0.0.1:8000/predict -H "Content-Type: application/json" -d '{"website_url": "https://text.npr.org/"}'

A few key benefits:

Deploy any pipeline or model: Agents, pipelines, RAG, chatbots, image models, video, speech, text, etc...
No MLOps glue: LitAPI lets you build full AI systems (multi-model, agent, RAG) in one place (more).
Instant setup: Connect models, DBs, and data in a few lines with setup() (more).
Optimized: autoscaling, GPU support, and fast inference included (more).
Deploy anywhere: self-host or one-click deploy with Lightning (more).
FastAPI for AI: Built on FastAPI but optimized for AI - 2× faster with AI-specific multi-worker handling (more).
Expert-friendly: Use vLLM, or build your own with full control over batching, caching, and logic (more).

⚠️ Not a vLLM or Ollama alternative out of the box. LitServe gives you lower-level flexibility to build what they do (and more) if you need it.

Here are examples of inference pipelines for common model types and use cases.

Toy model: Hello world LLMs: Llama 3.2, LLM Proxy server, Agent with tool use RAG: vLLM RAG (Llama 3.2), RAG API (LlamaIndex) NLP: Hugging face, BERT, Text embedding API Multimodal: OpenAI Clip, MiniCPM, Phi-3.5 Vision Instruct, Qwen2-VL, Pixtral Audio: Whisper, AudioCraft, StableAudio, Noise cancellation (DeepFilterNet) Vision: Stable diffusion 2, AuraFlow, Flux, Image Super Resolution (Aura SR), Background Removal, Control Stable Diffusion (ControlNet) Speech: Text-speech (XTTS V2), Parler-TTS Classical ML: Random forest, XGBoost Miscellaneous: Media conversion API (ffmpeg), PyTorch + TensorFlow in one API, LLM proxy server

Browse 100+ community-built templates

Self-host with full control, or deploy with Lightning AI in seconds with autoscaling, security, and 99.995% uptime.
Free tier included. No setup required. Run on your cloud

lightning deploy server.py --cloud

deploy.mp4

Feature Self Managed Fully Managed on Lightning

Docker-first deployment	✅ DIY	✅ One-click deploy
Cost	✅ Free (DIY)	✅ Generous free tier with pay as you go
Full control	✅	✅
Use any engine (vLLM, etc.)	✅	✅ vLLM, Ollama, LitServe, etc.
Own VPC	✅ (manual setup)	✅ Connect your own VPC
(2x)+ faster than plain FastAPI	✅	✅
Bring your own model	✅	✅
Build compound systems (1+ models)	✅	✅
GPU autoscaling	✅	✅
Batching	✅	✅
Streaming	✅	✅
Worker autoscaling	✅	✅
Serve all models: (LLMs, vision, etc.)	✅	✅
Supports PyTorch, JAX, TF, etc...	✅	✅
OpenAPI compliant	✅	✅
Open AI compatibility	✅	✅
MCP server support	✅	✅
Asynchronous	✅	✅
Authentication	❌ DIY	✅ Token, password, custom
GPUs	❌ DIY	✅ 8+ GPU types, H100s from $1.75
Load balancing	❌	✅ Built-in
Scale to zero (serverless)	❌	✅ No machine runs when idle
Autoscale up on demand	❌	✅ Auto scale up/down
Multi-node inference	❌	✅ Distribute across nodes
Use AWS/GCP credits	❌	✅ Use existing cloud commits
Versioning	❌	✅ Make and roll back releases
Enterprise-grade uptime (99.95%)	❌	✅ SLA-backed
SOC2 / HIPAA compliance	❌	✅ Certified & secure
Observability	❌	✅ Built-in, connect 3rd party tools
CI/CD ready	❌	✅ Lightning SDK
24/7 enterprise support	❌	✅ Dedicated support
Cost controls & audit logs	❌	✅ Budgets, breakdowns, logs
Debug on GPUs	❌	✅ Studio integration
20+ features	-	-

LitServe is designed for AI workloads. Specialized multi-worker handling delivers a minimum 2x speedup over FastAPI.

Additional features like batching and GPU autoscaling can drive performance well beyond 2x, scaling efficiently to handle more simultaneous requests than FastAPI and TorchServe.

Reproduce the full benchmarks here (higher is better).

These results are for image and text classification ML tasks. The performance relationships hold for other ML tasks (embedding, LLM serving, audio, segmentation, object detection, summarization etc...).

💡 Note on LLM serving: For high-performance LLM serving (like Ollama/vLLM), integrate vLLM with LitServe, use LitGPT, or build your custom vLLM-like server with LitServe. Optimizations like kv-caching, which can be done with LitServe, are needed to maximize LLM performance.

LitServe is a community project accepting contributions - Let's make the world's most advanced AI inference engine.

💬 Get help on Discord
📋 License: Apache 2.0

Read Entire Article