ATLAS

Overview

ATLAS (Adaptive Test-time Learning and Autonomous Specialization) is a local coding agent that brings frontier-style reasoning and verification to compact open models. It puts planning, candidate generation, quality scoring, sandboxed testing, and repair around the model so smaller models can tackle software work on local hardware without a hosted API or per-token fees.

Key Features

• Adds planning, candidate generation, quality scoring, sandboxed testing, and repair around compact open models.
• Uses a multi-phase V3 pipeline with structured planning, diverse sampling, budget allocation, repair, and refinement loops.
• Scores candidate quality with an energy-based Geometric Lens over the model's own embeddings, without an external oracle.
• Runs generated code in an isolated sandbox that supports multiple languages and can run generated and existing test suites.
• Provides local LLM inference through llama-server on consumer GPUs, with grammar-constrained decoding.
• Includes a native terminal UI, an atlas-proxy agent loop, and model-agnostic support for compatible GGUF models.

Use Cases

• Running coding assistance entirely on your own hardware without a hosted API or per-token fees.
• Helping compact open models handle real software work through planning, verification, and repair.
• Compiling, testing, and correcting generated code in an isolated execution environment before accepting it.
• Using compatible GGUF models across NVIDIA, AMD, Apple Silicon, Vulkan, or CPU-supported hardware.

Getting Started

• Use the one-shot bootstrap installer; it detects the Linux distribution and GPU vendor, installs the appropriate runtime, downloads model weights, builds the ASA steering vector, and starts the stack.
• After installation, run atlas in any project directory to launch the terminal UI.
• For a pinned or reviewable install, use the bootstrap script at a signed release tag or download and inspect the script before running it.

Deployment & Requirements

• GPU with 16 GB or more VRAM; NVIDIA CUDA is supported, AMD ROCm is community-tested, Apple Silicon Metal is supported through a native macOS hybrid path, and Vulkan is preview.
• The prebuilt CUDA image targets Blackwell RTX 50xx; older NVIDIA GPUs require a one-time local rebuild.
• Runtime: Docker, with nvidia-container-toolkit for NVIDIA and standalone Docker for AMD, or Podman.
• Python 3.9 or newer.
• About 20 GB disk for CUDA or about 22 GB for ROCm for model weights and container images.

Before You Adopt

• License: AGPL-3.0. Review its terms before using, modifying, or distributing the project.
• ATLAS is marked experimental, and the 3.1 line has known verification gaps; the source advises reviewing outputs and running your own tests until fixes planned for 3.2.0.
• No capability or reliability benchmark result is currently published for the release.
• In versions 3.1.0 to 3.1.7, a failing V3 candidate may still be written and reported as verified when no candidate passes tests.
• In versions 3.1.3 to 3.1.7, the lens reports calibration and active interventions but provides no usable signal with default settings.
• Sandbox commands have outbound network access by default.