Meta launches Muse Spark; Gemini 3.1 Pro still leads

Meta introduced Muse Spark in a company announcement this week, describing it as the company’s most capable model to date with improved reasoning and multimodal abilities across text, images and other inputs. Meta said the model is intended to handle more complex tasks and to support developers and enterprise users testing domain-specific applications.

Meta described technical changes to Muse Spark that include support for longer context windows, tighter integration of multimodal inputs and controls for safer, more predictable responses. The company plans to expand access through developer tools and product integrations so businesses and researchers can run the model on their own data and workflows.

Google’s Gemini 3.1 Pro continues to rank at or near the top on widely used public evaluation suites and independent tests. Those evaluations measure coding, reasoning and general knowledge tasks, including multi-task academic benchmarks, coding challenges and multi-turn dialogue assessments.

Recent rounds of third-party and public benchmarks show Muse Spark making gains in image understanding and contextual follow-through. Gemini 3.1 Pro retained an advantage on measures of coding accuracy, reasoning and general knowledge in many of the same evaluations.

Google has issued iterative updates to Gemini focused on latency, safety and specialized capabilities. Early enterprise customers using Gemini 3.1 Pro report high accuracy on reasoning and coding workloads and robustness on prompts that combine text and images.

Both companies reported safety features built into their models. Meta described filters and guardrails for Muse Spark intended to reduce harmful or misleading outputs and to monitor behavior during deployment. Google continues to refine safety layers and content filters for Gemini and to monitor real-world usage to guide updates.

Observers and developers tracking model releases note that relative rankings vary by the tasks and datasets used for evaluation. Vendors emphasize different strengths: Meta highlighted multimodal and context-handling improvements for Muse Spark, while Google emphasized Gemini 3.1 Pro’s top-tier performance across standard suites.

Organizations evaluating foundation models should compare task-specific performance, latency, cost, data privacy protections and safety controls when testing providers and selecting models for production use.