Descriptions:
Fahd Mirza puts Meta’s newly released Muse Code terminal coding agent through a hands-on evaluation on Ubuntu, powered by the Muse Spark 1.2 model. Muse Code is designed for complex repository-scale work: it maintains persistent agent state across a full session, fans out large tasks to parallel isolated worker processes to avoid file conflicts, and logs all actions locally for crash recovery and resumption.
Mirza runs two distinct tests. First, a broken FastAPI and Docker-based cabin booking application with five deliberately planted bugs — Muse Code identifies and fixes all five, verified visually in the browser. Second, a creative challenge: a single-file HTML simulation of a global railway convergence hub with ten trains from systems including the Trans-Siberian, Tokyo, New York, and Indonesian networks, all required to run without collisions. Notably, Muse Spark 1.2 takes screenshots in headless Firefox mode to visually validate its own output — a multimodal self-verification loop that Mirza flags as a differentiator.
On the performance side, Meta benchmarks Muse Spark 1.2 just behind Claude Opus 5 on terminal and real-world software engineering tasks. Pricing on the contributor tier is $0.10 per million input tokens and $0.20 per million output tokens, with a one-million-token context window. Mirza credits DeepSeek’s pricing pressure for pushing frontier models to this price point, and notes the standard privacy-preserving tier carries a higher cost. The model is available via API at dev.meta.ai.
📺 Source: Fahd Mirza · Published August 05, 2026
🏷️ Format: Hands On Build
