Meta unveils Muse Spark (Avocado), pushing AI from chat to action with native multimodal reasoning

What happened

Meta has released Muse Spark — code-named Avocado — a new “native multimodal reasoning” model that the company frames not as a better chatbot but as the foundation for an “action system.” Meta says Muse Spark can process text, images and environmental signals, call external tools, and orchestrate multiple agents to solve complex tasks. It has been reported that Alexandr Wang — the engineer widely discussed in media reports as having been recruited to Meta — contributed to the project after a nine‑month quiet period; those recruitment details and the cited $14.3 billion figure should be treated as unverified reports.

Key technical claims

Meta says the launch reflects a full stack rebuild: model architecture, training pipeline, data management and new infrastructure (including a next‑generation data center called Hyperion). The company claims Muse Spark reaches comparable performance to earlier large models such as Llama 4 Maverick while using an order of magnitude less pretraining compute (FLOPs). Meta also introduced a “visual chain‑of‑thought” to expose and compress reasoning steps, plus a multi‑agent “thinking mode” that coordinates parallel reasoning — all designed to raise token efficiency and make large‑scale test‑time reasoning practical.

Why this matters

Why should Western readers care? Muse Spark signals a strategic shift from standalone model capability to system‑level AI: model + tools + environment + multi‑agent orchestration. That matters for product design and for competition with Chinese firms such as Baidu (百度), Alibaba (阿里巴巴) and Tencent (腾讯), which are pursuing their own multimodal and agentic systems. It also intersects with geopolitics: export controls on high‑end chips and growing scrutiny of advanced AI systems mean that engineering efficiency (doing more with less compute) is becoming as important as raw scale.

Safety and reception

Meta says it ran system‑level safety evaluations under an “Advanced AI Scaling Framework,” including domain‑specific tests in health and biology and collaboration with more than 1,000 doctors for medical training data. Early reactions in the tech community have been enthusiastic, with prominent researchers and industry figures congratulating the team. But independent benchmarking and broader safety audits will be needed before Muse Spark’s real‑world capabilities and risks are fully understood.