Meta AI is Competitive with Mus Spark AI

by

Meta Muse Spark (originally codenamed Avocado) is Meta’s first model from its new Meta Superintelligence Labs (MSL). It launched on April 8, 2026 and will be the first of a new Muse family of models.

Meta describes it as a natively multimodal reasoning model designed from the ground up for vision + language integration (not just stitched-on vision like earlier Llama models).

It natively supports:

Tool use
Visual chain-of-thought reasoning
Multi-agent orchestration

It is not being positioned as Meta’s absolute flagship best ever model, but rather as an efficient, competitive foundation that prioritizes predictable scaling, lower compute costs, and practical performance for real-world tasks (multimodal perception, reasoning, health, agentic workflows).

Core Innovation: Multi-Agent Orchestration

The old bottleneck in AI was forcing one model to handle reading → planning → tool calls → final answer in a single sequential stream. Muse Spark’s biggest leap is multi-agent orchestration:

Multiple copies of the model work on the same problem in parallel.
They compare, debate, or merge results (closer to a small team than a solo assistant).
This is called Contemplating mode (rolling out gradually).

It delivers big gains on hard benchmarks without the latency penalty of “single-agent thinking longer”:

Humanity’s Last Exam: 58% (in Contemplating mode)
FrontierScience Research: 38% (in Contemplating mode)

This shifts the scaling story — better performance no longer comes only from making one model bigger; you can spend compute more intelligently at runtime by adding parallel agents.

Three Scaling Axes
Muse Spark was built on a completely rebuilt AI stack with three deliberate axes:

Stronger pretraining — Better architecture, optimization, and data curation for world/code understanding.
Steadier reinforcement learning (RL) — Improves answer quality post-pretraining.
Test-time reasoning — Extra compute is spent only when a problem is hard (via thinking-time penalties + multi-agent orchestration).

Meta claims the new pretraining recipe reaches similar capability with over 10× less training compute than Llama 4 Maverick (its previous mid-size flagship)