---
title: README
emoji: 🏢
colorFrom: yellow
colorTo: indigo
sdk: static
pinned: false
---

# **TheStage AI Platform**

Inference optimization for LLMs, diffusion, and voice. Self-hosted or cloud. Works on NVIDIA GPUs, Apple Silicon, and edge devices.

**Links:**

[Web App](https://app.thestage.ai/) • [Docs](https://docs.thestage.ai/) • [Hugging Face](https://huggingface.co/TheStageAI) • [X](https://x.com/TheStageAI) • [LinkedIn](https://www.linkedin.com/company/thestageai) • [Discord](mailto:sergey@thestage.ai) (request invite) • [Email](mailto:support@thestage.ai)

---

# **What is TheStage AI**

TheStage AI is an inference optimization stack. It helps you compress, compile, and serve models. You keep control of the accuracy versus performance trade-off.

---

# **Products / Components**

- [**ANNA (Automatic Neural Network Acceleration)**](https://docs.thestage.ai/qlip/docs/source/anna_api.html)
    
    Automated compression analysis under user-defined constraints (size, MACs, latency, memory). Outputs a QlipConfig for compile and serve.
    
- [**Qlip**](https://docs.thestage.ai/qlip/docs/source/index.html)
    
    Full-stack optimization and inference framework. Quantization, sparsification, and compilation for NVIDIA GPUs (Apple Silicon supported). Produces pre-compiled (non-JIT) artifacts with dynamic shapes and mixed precision. Triton-based serving.
    
- [**Elastic Models**](https://docs.thestage.ai/tutorials/source/elastic_transformers.html)
    
    Qlip-optimized models with S / M / L / XL performance tiers (availability varies). L/M/S may include quantization or pruning for faster inference.
    
- [**TheStage CLI**](https://docs.thestage.ai/platform/src/thestage-ai-cli.html)
    
    Manage projects, tokens, and hardware from the terminal. Launch/monitor jobs, rent instances, and stream logs.
    
- [**TheStage Platform**](https://app.thestage.ai/)
    
    Web UI and APIs for instances, models, and deployments. Includes the [**Playground**](https://app.thestage.ai/) to test Elastic Models, switch hardware, and compare tiers before deployment.
    

---

# **Key features**

- **Elastic Models with S/M/L/XL tiers per model** (choose cost, quality, and memory balance; availability varies).
- **ANNA constraint-driven compression analysis** (outputs a QlipConfig for compile and serve).
- **Qlip compiler and runtime** (pre-compiled engines; no runtime JIT; dynamic shapes; mixed precision).
- **OpenAI-compatible HTTP serving** (deploy and scale models through a standard API).
- **Playground to test models and hardware** (compare performance and tiers before deployment).
- **Self-host or run in the cloud** (use your own infrastructure; keep data private).
- **Hardware support: NVIDIA (incl. Jetson), Apple Silicon, and edge targets** (NPUs, DSPs, and MCUs per model).
- **Comprehensive tutorials and documentation** (from setup to evaluation and production).

---

# **Quickstart**

- Install CLI: `pip install thestage`
- Set token: `thestage config set --api-token <YOUR_API_TOKEN>` (get it in the web app)
- Use `elastic_models` in your code and choose a tier (S/M/L/XL). See Markdown version for a snippet.
- Diffusion and voice examples are in the docs.

---

# **Serving**

OpenAI-compatible API flow with Modal is documented (single- and multi-GPU).

Start here: https://docs.thestage.ai/

---

# **Supported hardware**

- NVIDIA GPUs (incl. Jetson where applicable)
- Apple Silicon
- Edge/embedded devices

---