Meta released Muse Spark 1.1, a model that beats Zhipu's GLM-5.2 on coding benchmarks while undercutting it on price. The model scored 71.3 on coding tasks compared to GLM-5.2's lower performance, at a cost of $0.26 per task.
Performance gains accelerated rapidly. Over three months, Muse Spark 1.1 climbed eight points on the Artificial Analysis Intelligence Index to reach a score of 51. The hallucination rate fell sharply from 73 percent to 38 percent, a critical improvement for production systems where false outputs create costly errors.
The cost advantage matters for deployment at scale. At $0.26 per task, the model undercuts alternatives while delivering superior coding capabilities. This positions it as a practical choice for enterprises building AI-powered development tools or automating code generation workflows.
Coding performance has become a key battleground in the large language model race. Strong code generation attracts developers and tools companies, driving both adoption and revenue. Muse Spark 1.1's edge over GLM-5.2 signals Meta's continued investment in closing the gap with frontier models from OpenAI and Anthropic.
The hallucination reduction speaks to improvements in training or inference techniques. Cutting hallucinations in half changes the model's real-world viability. Code generation systems especially need reliability, since fabricated function names or logic errors compound into broken deployments.
Context matters here. Meta has been aggressive in open-source model releases and proprietary applications. Muse Spark targets the commercial coding assistant market where companies like GitHub Copilot and Claude already have footholds. The benchmarks suggest Meta has a competitive offering.
The trajectory indicates Meta continues refining its models at speed. An eight-point