Meta Muse Spark

Muse Spark is the first model from Meta Superintelligence Labs (MSL), the new research division led by Alexandr Wang. Internally codenamed "Avocado," it represents a clean-sheet architecture rebuilt from scratch over nine months — this is not a Llama derivative. MSL deliberately abandoned the Llama lineage to explore a fundamentally different approach to multimodal intelligence, one that natively fuses text, image generation, and speech synthesis into a single model rather than bolting modalities together post-training.

The most distinctive technical innovation is "thought compression," a reinforcement learning technique that trains the model to reach correct answers using fewer reasoning tokens. Where GPT-5.4 consumed 120M tokens and Claude used 157M tokens across the full evaluation suite, Muse Spark completed the same benchmarks with just 58M tokens — roughly 2-3x more efficient. This matters enormously for deployment at Meta's scale, where the model serves Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta smart glasses simultaneously.

Muse Spark operates in three distinct modes: Instant (fast responses, no extended reasoning), Thinking (single-agent chain-of-thought), and Contemplating (multi-agent parallel reasoning where several reasoning threads explore different solution paths concurrently). The Contemplating mode is particularly novel — it orchestrates multiple reasoning agents in parallel, then synthesizes their findings, achieving strong results on complex problems like HealthBench Hard (42.8%, #1 among all models) and CharXiv Reasoning (86.4%, also #1).

Notably, Muse Spark is entirely closed-source and consumer-only via meta.ai — a major departure from Meta's open-source tradition with Llama. There is no public API, no downloadable weights, and no pricing structure. This strategic pivot signals MSL's focus on direct consumer experiences over developer ecosystem building.