Meta's Muse Spark Open Targets Agentic Coding

Meta AI has introduced Muse Spark Open, an open-weight model aimed squarely at agentic coding, according to the company's research blog. The release leans on a 1-million-token context window, a scale meant to keep entire codebases, logs, and task histories in view as the model works through multi-step engineering problems.

The pitch here is less about raw chat quality and more about sustained, tool-using work. Agentic coding — where a model plans, edits across many files, runs through long sequences of actions, and course-corrects — tends to strain shorter context limits. A seven-figure token budget gives the model room to reason over large repositories without aggressive truncation or retrieval gymnastics.

Why it matters

Meta is framing Muse Spark Open as the most capable US open-weight model to date, a claim that signals where it wants to compete as open releases from other regions have gained ground. For developers and teams, the practical appeal is concrete:

• Open weights under the Llama license, allowing self-hosting and customization
• A 1M-token context suited to whole-project reasoning
• A design orientation toward agentic, long-horizon code tasks rather than single-shot completions

As with any Llama-licensed release, the terms carry use restrictions that fall short of a fully permissive open-source license, so teams will want to check the conditions before building on it. Still, a large-context, coding-focused model with published weights lowers the barrier for groups that need to run agents in controlled environments. The full details are available on Meta's research blog.