Meta announces Muse Spark, a multimodal model that achieves comparable capabilities to Llama 4 Maverick with over an order of magnitude reduction in pretraining compute through architectural and optimization improvements. The model's scaling is tracked across three axes: pretraining, reinforcement learning, and test-time reasoning.
Models
Muse Spark: Scaling towards personal superintelligence
Meta's Muse Spark achieves Llama 4 Maverick-level multimodal performance with over 10x less pretraining compute through architectural and optimization breakthroughs.
Wednesday, April 8, 2026 12:00 PM UTC2 MIN READSOURCE: Hacker NewsBY sys://pipeline
Tags
models
/// RELATED
StrategyApr 8
Meta debuts the Muse Spark model in a ‘ground-up overhaul’ of its AI
Meta launches Muse Spark with parallel agents and reasoning capabilities, part of a $14.3B restructuring under new Superintelligence Labs to narrow the gap with OpenAI and Anthropic.
ModelsApr 8
Meta unveils Muse Spark, its first AI model since hiring Alexandr Wang and a bellwether for CEO Mark Zuckerberg’s multibillion-dollar AI push
Meta shifts from open-weight Llama to closed proprietary AI with Muse Spark, signaling Zuckerberg's bet on competing directly with OpenAI and Anthropic rather than commoditizing AI via open-source.