inferstep/ATLAS

Adaptive Test-time Learning and Autonomous Specialization

What it solves

ATLAS is a local coding agent designed to enable compact open-source models to perform complex software engineering tasks. It solves the problem of smaller models lacking the native reasoning and verification capabilities of frontier models by surrounding the LLM with a system of planning, candidate generation, quality scoring, and sandboxed testing.

How it works

ATLAS implements a multi-layered architecture to enhance model performance:

  • V3 Pipeline: A multi-phase generation process that uses constraint-driven planning (PlanSearch), diverse candidate sampling (DivSampling), and iterative repair loops (PR-CoT) to turn a single prompt into a verified code candidate.
  • Geometric Lens: An energy-based scoring system that uses the model's own embeddings to predict candidate quality via an MLP and XGBoost ensemble, allowing the system to veto or select candidates without an external oracle.
  • Atlas-Proxy: A Go-based orchestration layer that handles tool-call routing, enforces output grammar via GBNF schemas, and applies bias mitigations to steer the model toward structural edits.
  • Sandbox: An isolated execution environment where generated code is compiled, linted, and tested against existing or self-generated test suites before being accepted.
  • Local Inference: Uses llama-server for GPU-accelerated quantized inference, supporting NVIDIA, AMD, Apple Silicon, and Vulkan.

Who it’s for

It is for developers who want a powerful, self-hosted coding agent that runs entirely on their own hardware (16GB+ VRAM) to maintain privacy, avoid per-token API fees, and use open-weight models.

Highlights

  • Local-First: No hosted APIs or third-party keys required; runs entirely on consumer GPUs.
  • Verification Loop: Code is verified in a sandbox before being written to the project, reducing hallucinations.
  • Agentic Workflow: Includes a native Bubbletea TUI with slash commands for file context and shell execution.
  • Model Agnostic: Supports various GGUF models and provides tools for local Lens and ASA training/retraining.
  • Hardware Versatility: Broad support for CUDA, ROCm, Metal, and Vulkan.

Related

  • Project
    facebookresearch/atlas-lean

    ATLAS is a research project that uses large language models to automatically translate textbook mathematics into formal Lean 4 libraries, providing reusable, machine‑checkable mathematics for theorem proving and AI research.

  • Project
    styler-ai/ProjectAtlas

    ProjectAtlas is a Rust-native CLI and MCP server that creates a local, intelligent map of your codebase to help AI coding agents navigate large repositories more efficiently and reduce token costs.

  • Project
    Avarok-Cybersecurity/atlas

    A pure Rust and CUDA LLM inference engine that provides a lightweight, high-throughput server for large models without requiring Python or PyTorch.

  • Project
    esengine/DeepSeek-Reasonix

    Reasonix is an open‑source, locally‑run AI coding assistant that can read, edit, and execute code in your project. You configure any OpenAI‑compatible model (DeepSeek preset included) and interact via a desktop app, terminal, web UI, or VS Code extension. Features include permission levels, file‑level rewind, plugin‑driven tools, multi‑model composition, and a single‑binary distribution. Install via platform‑specific installers (2.x Studio) or npm (1.x CLI), then describe tasks in plain language and let the assistant modify your code under your supervision.

  • Project
    atlassdk/atlas

    Atlas is a decentralized coordination layer that allows AI agents across different blockchain ecosystems to discover each other, negotiate tasks, and settle payments.