Benchmark: Meta's open-weight Muse Spark Open fails as a coding agent

Meta's open-weight Muse Spark Open (816B parameters, 42B active) scored 0.091 and 0.0877 out of 1 on the agentic coding Forge 2.0 test, versus 0.8523 for the best build. Its Gauntlet 7.2 attempt got no verdict after a provider timeout. Local Gemma 4, Granite 4.2, LFM2.5, Xing 4.0 and Nemotron 3.5 Bolt did no better.

• Muse Spark Open: 816B parameters, 42B active, released 15 October
• Forge 2.0: 0.091 and 0.0877, 53rd and 54th of 64
• Gauntlet 7.2: provider timeout after 13 calls, $0.19
• Qwen3.8-27B on a MacBook scored 0.2576, nearly three times higher