Meta Unveils Muse Spark As Its First Proprietary AI Model

San Francisco: Meta has launched Muse Spark, its first proprietary AI model and the inaugural release from its newly formed Meta Superintelligence Labs. The launch marks a sharp break from Meta’s open-source identity and its most serious attempt yet to compete with OpenAI, Anthropic, and Google.

Muse Spark was built from scratch over nine months. Meta did not iterate on its Llama models. It rebuilt its entire AI stack, architecture, infrastructure, and data pipelines, from the ground up. The result is a model Meta claims achieves comparable capability to Llama 4 Maverick using one tenth of the compute.

The model is led by Alexandr Wang, Meta’s chief AI officer, who joined the company nine months ago as part of a USD 14.3 billion deal for a 49% stake in Scale AI. Wang heads Meta Superintelligence Labs, built from researchers recruited from OpenAI, Anthropic, and Google. Muse Spark, internally codenamed Avocado, is the first tangible output of that investment.

On benchmarks, Muse Spark scored 52 on the Artificial Analysis Intelligence Index. It ranks fifth overall, behind GPT-5.4, Gemini 3.1 Pro, and Claude Opus 4.6. Its clearest lead is in health. On HealthBench Hard it scored 42.8 versus GPT-5.4’s 40.1.

Contemplating mode runs multiple AI agents in parallel rather than one at a time. It scored 50.2% on Humanity’s Last Exam. GPT-5.4 Pro scored 43.9% on the same test. Meta has openly acknowledged gaps in coding and agentic tasks.

The model accepts voice, text, and image inputs. Output is text only. Three modes cover different use cases: Instant for quick queries, Thinking for complex reasoning, and Contemplating for multi-agent tasks. A Shopping Mode layers user interest and behaviour data on top of the model. That signals how Meta plans to monetise differently from rivals.

Muse Spark is live now on meta.ai and the Meta AI app, free to use. It will roll out across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta glasses in the coming weeks. Meta said it plans to spend USD 115 to USD 135 billion on AI infrastructure in 2026. Following the announcement, Meta’s stock rose about 9%.

The biggest change is Meta’s move away from open-source AI. For three years, Llama models defined the developer ecosystem. Developers built businesses and tools on its open weights. Muse Spark ends that era.
Meta is not sharing the model’s weights publicly. It says it hopes to open-source future versions. No timeline has been given. The developer community has noticed and is watching closely.