Meta Muse Spark 1.2 Honest Review: Loses the Benchmarks, Wins as an Agent

Meta just shipped two things at once: Muse Spark 1.2, the model, and Muse Code, an agentic harness that wraps around it. The interesting part is that they pull in opposite directions. On raw coding the model is a clear underdog. As an agent it is suddenly near the top. Here is what actually held up when I tested it. Prefer the quick version? I covered the benchmark split, the broken games, the harness features, and a head to head against Qwen 3.8 in this video: What Muse Spark and Muse Code are Muse Spark 1.2 is Meta's new coding model. Muse Code is the harness you actually run it in. The harness keeps a local log of every tool call and every edit, so if a run crashes it picks up exactly where it left off instead of starting over. It ships with built-in skills too: /plan turns a task into an approval-gated plan, /grill stress-tests that plan before you run it, and /goal keeps the agent pushing toward the objective. Where it loses On the coding benchmarks it never comes first. Second on Terminal Bench behind Opus 5, third on DeepSWE, and behind Opus 5 on Meta's own internal benchmark. The games I had it build showed the same weakness. They were laggy and half-finished, one dragon bounced on the spot, one character's arms were missing, another spun around when you tried to walk backwards. On the same prompts, Qwen 3.8 and Fable 5 built noticeably cleaner, more playable versions. Where it wins Point the same model at tools inside Muse Code and the picture flips. On agent and tool-use benchmarks it jumps to first. The crash-resume log is the standout feature for anyone running long agent jobs. To test it properly I gave both Muse Spark and Qwen 3.8 the same task: read a guide and turn it into a reusable skill. Muse Spark replied faster and produced the sharper result. It analyzed the guide, built a proper table, and got specific instead of generic. Who should run it If you want the best raw coding model, Opus 5 still wins and Qwen 3.8 still builds cleaner. If you care about agent workflows, tool use, and not losing progress when a long run dies, Muse Code is worth a serious look even though the model underneath loses the benchmark race. My verdict The model is not the story. The harness is. Muse Spark 1.2 is a reminder that in 2026 the wrapper around a model can matter as much as the weights. Have you tried Muse Spark or Muse Code yet? Tell me in the comments. If you like quick, honest AI-tool reviews, follow my YouTube channel for the next one. Tags: #ai #metaai #musespark #aiagents #coding``