ATLAS
A.T.L.A.S. ('Adaptive Test-time Learning and Autonomous Specialization') by Inferstep is an open-source, local-first terminal coding agent (AGPL-3.0) designed to bring frontier-grade reasoning to compact open-weight models running on consumer hardware.
Addressing the reality that raw 7B–14B models struggle with complex multi-file software engineering tasks when relying on single-shot completions, ATLAS surrounds the model with an intelligent system scaffolding. It implements test-time compute scaling: dynamically decomposing goals into hierarchical plans, generating multiple candidate code solutions, ranking candidates via quality scoring, compiling and executing code inside isolated sandboxes, and automatically repairing test failures. Compatible with GGUF models via llama.cpp, ATLAS runs locally across Apple Silicon, NVIDIA, AMD, and CPUs with zero cloud API fees.
For privacy-conscious developers and offline engineering teams, ATLAS unlocks remarkable code generation capabilities without external token billing. However, allocating substantial test-time compute (generating multiple candidates and running local test suites) multiplies local inference duration, smaller compact models still face reasoning limits on high-level architecture, and the codebase carries copyleft AGPL-3.0 licensing.
Features
Why choose it
Trade-offs
Maintenance check
- Last commit
- 2026-10-24
- Commits, last 90 days
- 55
- Releases, last 90 days
- 4
- Latest release
- v3.1.7 (2026-10-25)
- Open issues
- 147
- Recent issues with at least one comment
- 0/1
Oldest week on the left; the latest four weeks are orange. Source: GitHub commit activity.
Status rule: archived on GitHub = Archived; last commit within 90 days = Active; 91 to 365 days = Slowing; older = Stale. All figures come from the GitHub API on the date shown.
Most discussed open issues
- #35 JetBrains plugin (the VS Code extension is on dev, for 3.2.0)16 comments · 2026-04-13
- #245 CI: run each service's Python tests on the Python its image uses6 comments · 2026-09-28
- #238 Held-out evaluation suite: frozen tasks and graders with paired controls6 comments · 2026-09-28
- #215 Content loops are detected late and use more than half of a session6 comments · 2026-09-28
- #202 OpenSSF Scorecard: pin the remaining installs and re-check after 3.2.06 comments · 2026-09-28
Issue titles are quoted from GitHub. Open the link to check whether a problem still applies to the version you use.
Use cases
Alternatives
Comments
No comments yet. Be the first!