Parallax
Labs

Creating amazing AI models that are highly capable, yet tiny, with a focus on reasoning capabilities as well as coding.

Try Parallax 2 Now

Introducing Parallax 2

Our flagship reasoning model, with 9 billion parameters. Trained on over 202,000 elite reasoning, coding, and agentic traces from models like Claude Opus and GPT-5.6. It is comparable to frontier models, but Parallax 2 can run all locally. Experience top-tier reasoning locally with a 262K context window.

Available Now

Introducing Parallax App

Our simple, yet powerful local-first AI workbench. Chat with Parallax 2, build complex projects with Parallax 2 Agent, and chat with Parallax 2 in Chat Mode. No cloud, no APIs, no subscriptions. Just you, your hardware, and the most capable local AI suite ever built.

Available Soon

Parallax 2 is Officially Available

We are thrilled to announce the public release of Parallax 2. After a successful training run on over 202,456 frontier traces, our flagship 9-billion-parameter model is ready for the world. It is designed to be a top-tier local AI for reasoning, coding, and autonomous agentic tasks.

Parallax 2 is now available for download on HuggingFace. You can run it locally using tools like LM Studio or Ollama, and experience frontier-level intelligence on your own hardware.

A Game Built by AI: Parallax 2 on an RTX 5070

In a recent internal test, we tasked Parallax 2 with building a multiplayer, 3D game from scratch. Running locally on an RTX 5070, the model autonomously architected a multi-file project (JS, CSS, HTML), wrote a WebRTC-based networking layer, and generated a full 3D game engine using Three.js.

The model's ability to build a complex codebase and debug its own errors in real-time demonstrates the power of our agentic training data. This is a new era for local AI software engineering.

The $32 Frontier: How We Trained a 9B Model on a Budget

While cloud giants spend billions on data centers, we rented a single RTX PRO 6000 for $32. By curating elite traces from Opus, Fable, and OpenSWE, we distilled frontier-level intelligence into a 9B model that runs on consumer hardware.