Meta’s Muse Spark Open Scores 39, Leads U.S. Open-Weight Models

Meta’s Muse Spark Open model scored 39 on the Intelligence Index, putting it ahead of NVIDIA’s Nemotron 3.5 Super VL among U.S. open-weight models while leaving three Chinese models ahead or level globally.

Xiaomi’s MiMo-V2.6-Pro scored 46, Zhipu’s GLM-5.3 scored 45 and DeepSeek’s V4.1 Flash tied Muse Spark Open at 39. NVIDIA’s model scored 26, giving Meta’s model a 13-point lead over its U.S. competitor.

Results varied across test types. Among the nine highest-scoring open-weight models, Muse Spark Open ranked last across four agent tests. It ranked fourth across the other six tests. The entire six-point gap between Muse Spark Open and GLM-5.3 came from the agent tests.

Muse Spark Open scored 25 on the AA-Omniscience knowledge-accuracy test, the highest result among open-weight models. Its hallucination rate was 24%, the lowest among open-weight models with an overall score above 30.

The model averaged about 40,000 output tokens per test, 43% fewer than GLM-5.3. With API prices close to each other, Muse Spark Open’s average cost was $0.84 per test: less than half GLM-5.3’s and about three times DeepSeek V4.1 Flash’s.

Meta’s closed flagship Muse Spark 1.2 scored 40, one point above Muse Spark Open. Muse Spark 1.2 was introduced in August.