Meta rolled out Muse Spark 1.1 on Thursday, positioning the new multimodal model as a leap forward for agentic AI. The company described the upgrade as a “significant” step from its predecessor, Muse Spark 1.0, emphasizing gains in computer use, coding, and the ability to manage complex, multi‑stage projects.
Superintelligence Labs, the research arm that built the model, is led by Alexandr Wang, who previously headed Scale AI. Spark 1.1 will be accessible through a public preview of the Meta Model API, and it is already live in the Thinking mode of the Meta AI app and on meta.ai.
The model’s core strength lies in its agentic capabilities. Spark 1.1 can orchestrate a hierarchy of agents: a main planner creates a roadmap, then delegates subtasks to parallel sub‑agents that execute, retrieve information, and report back. This structure lets the system optimize end‑to‑end latency and keep track of up to one million tokens of context, compressing and recalling critical steps as needed.
Beyond planning, Spark 1.1 can take direct actions on a user’s computer. Meta says the model can navigate unfamiliar interfaces, maintain context across long sessions, and decide when to automate versus when to act manually. The feature is optional and requires user consent, but it promises to handle routine tasks with minimal interaction.
On the coding front, the model shows notable improvements. It can diagnose bugs, add features to enterprise‑grade systems, and manage large‑scale code migrations, especially in complex database environments. Developers testing the preview have reported faster turnaround on intricate programming challenges.
Multimodal input is another highlight. Spark 1.1 can process images and audio, generating ultra‑descriptive captions and using visual or auditory cues to inform its actions. The capability mirrors experiences seen in other AI products, such as Gemini Live’s camera mode, but adds the layer of agentic decision‑making.
Meta also underscored the model’s safety architecture. Built on the Advanced AI Scaling Framework, Spark 1.1 reportedly resists jailbreak attempts, prompt injection, and other common attacks. The company said extensive evaluations were conducted before release.
The announcement arrives as competitors like OpenAI and Anthropic continue to launch advanced models, signaling Meta’s renewed push into the AI race after a year of internal reorganization. By bundling agentic functions, multimodal perception, and robust safety, Meta aims to position Muse Spark 1.1 as a versatile tool for developers and power users alike.
Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.