Meta launches Muse Spark, its first post-Llama model - hosted only, no open weights

Meta released Muse Spark, a hosted model with no open weights - a sharp departure from the Llama line. Benchmarks put it roughly on par with Opus 4.6, Gemini 3.1 Pro, and GPT 5.4, though Meta acknowledges weaknesses in agentic coding tasks. The model ships in two modes on meta.ai (Instant and Thinking), with a slower “Contemplating” mode promised later.

The more interesting story is the tooling wired into Meta’s chat interface. Simon Willison extracted descriptions of 16 tools, including web browsing, a Python sandbox (Code Interpreter-style with Python 3.9), HTML/SVG artifact rendering, image generation, sub-agent spawning, and a visual grounding tool that performs object detection with bounding boxes, points, or counts on generated or uploaded images. There’s also a semantic search tool that queries Instagram, Threads, and Facebook posts - scoped to content visible to the logged-in user - with filters for author, celebrity mentions, and engagement signals.

Notably, the sandbox lets users chain tools together: generate an image, analyze it with OpenCV, then run visual grounding on the result. Meta didn’t lock down tool metadata behind prompt guards, so the full capability surface was discoverable without jailbreaks. The platform also supports third-party account linking for Google and Outlook calendars and email.