Ante: Antigma's 15MB Rust agent hits 82.7% on Terminal-Bench

Antigma Labs pushed source for Ante, a coding agent that ships as a single Rust binary of about 15MB with grep, git, and a pinned build of llama.cpp all embedded in the same process, according to the project's GitHub README. The pitch is a self-contained alternative to Claude Code or Codex that can run local GGUF models with zero runtime dependencies and no Antigma account required.

The headline number is 82.7% on Terminal-Bench 2.1, across 89 tasks at five trials each, with 368 of 445 trials passing and DeepSeek V4 Flash 0731 as the underlying model. Antigma also reports that on twenty parallel tasks Ante consumes roughly seven times less peak memory, nine times less average CPU, and five times less disk I/O than Claude Code. Source is Apache 2.0; the prebuilt binary sits under a separate 'Binary Preview Terms' license during alpha, with commercial use permitted.

Two things are worth reading twice. First, that Terminal-Bench score is tethered to DeepSeek's model as much as to Ante's harness, so a different backend would produce a different number. Second, the resource comparison is Antigma's own measurement rather than a third-party benchmark, and the README does not describe the workload behind the twenty parallel tasks or how the same protocol scores Claude Code on Terminal-Bench 2.1. The specifics of the Binary Preview Terms are also named but not spelled out.

For teams that pay by the CPU-second on CI, or that run coding agents inside regulated environments where a hosted API is a non-starter, the value proposition writes itself: a static Rust binary is easy to distribute and the local GGUF path keeps the whole loop on your own hardware. It is also the second Rust-first agent project we have tracked this week, after Cloudflare's Kitesurf browser for agents, which suggests the systems-language crowd has decided coding agents should not stay in Node forever.