Muse Spark Unveiled

Meta has unveiled Muse Spark. It is a new model coming almost exactly 1 year after Llama 4, and it is currently offered only in hosted form, not as open weights. The API is also still at a private API preview stage accessible only to some users, and general users can try it via Facebook or Instagram login on meta.ai.

In its own benchmarks, it presented figures competitive with Opus 4.6, Gemini 3.1 Pro, and GPT 5.4, but it lagged behind on Terminal-Bench 2.0. Meta stated it will continue investing especially in areas where the current performance gap is large, such as long-horizon agentic systems and coding workflows.

The meta.ai chat is exposed in two modes, Instant and Thinking. A Contemplating mode that uses longer reasoning time will be added later, aiming for behavior closer to the more deeply-thinking Gemini Deep Think or GPT-5.4 Pro. In actual experiments, Instant produced a somewhat simple SVG, while Thinking produced better results along with an HTML shell, and an unused Playables SDK v1.0.0 library was also visible on the Thinking side.

When asked about tools, the chat revealed the names, parameters, and descriptions of 16 tools in fairly great detail. The key ones are as follows.

• browser.search / browser.open / browser.find: web search, loading result pages, pattern search
• meta_1p.content_search: semantic search of Instagram, Threads, Facebook posts
• meta_1p.meta_catalog_search: search of the Meta product catalog
• media.image_gen: image generation in artistic/realistic modes
• container.python_execution: Python 3.9 sandbox execution
• container.create_web_artifact: creation of HTML or SVG artifacts
• container.file_search / view / insert / str_replace: file search and editing
• container.visual_grounding: finding objects in images and returning them as bbox, point, count
• subagents.spawn_agent: running an independent sub-agent
• third_party.link_third_party_account: integration with Google Calendar, Outlook Calendar, Gmail, Outlook

The visual_grounding experiment was particularly strong. On a raccoon photo, it pinpointed a coffee cup, banana peel, newspaper, eyes, nose, the raccoon's body, and a trash can lid with points and bounding boxes, and in count mode it even counted raccoon whisker and raccoon paw claw. Thanks to this, Meta AI gives the impression of a fairly full-fledged agent harness that ties together image generation, Python analysis, object recognition, and sub-agent invocation all in one place.