Ante 0.2 Ships a 15MB Local Coding Agent Managing llama.cpp Offline

Exciting-Camera3226 · reddit · 2026-08-10

The dev team released Ante 0.2, a 15MB local coding agent. Its core highlight is fully offline capability, managing the llama.cpp inference engine so users can run the entire agent loop by simply pointing it to a local GGUF model file.

Key features include:

Automated hardware adaptation: Installs and verifies the matched llama.cpp build (supporting Apple Metal, CUDA, Vulkan, etc.).
Smart discovery & memory estimation: Scans local disks for model files and estimates RAM/VRAM requirements before loading.
Seamless switching: Local models sit alongside hosted ones as ordinary providers, allowing users to flip to local for sensitive repos and back to frontier models for complex tasks.

Regarding performance, the team is transparent about the gap: Qwen3.6 27B scores 56.2% on Terminal-Bench 2.1. The tool ships as a single self-contained binary and has processed nearly 7 trillion tokens since its preview launch.

Original post →

Ante 0.2 Ships a 15MB Local Coding Agent Managing llama.cpp Offline

More from coding & agent