Muse Spark — first model from Meta Superintelligence Labs (MSL)

TL;DR

Alexandr Wang announces Muse Spark, the first model from Meta Superintelligence Labs (MSL) and now powering Meta AI. Described as a natively multimodal reasoning model with tool use, visual chain-of-thought, and multi-agent orchestration, built on a from-scratch rebuild of Meta’s AI stack over the preceding nine months. The launch includes a separate “contemplating mode” that orchestrates multiple agents reasoning in parallel, pitched as competitive with Gemini Deep Think and GPT Pro. Private API preview is open today to select partners, with future versions planned to be open-sourced.

Key claims

Muse Spark is positioned as the most powerful model Meta has released and is the result of a nine-month rebuild of infrastructure, architecture, and data pipelines [tweet 1].
The model is natively multimodal with tool-use, visual chain-of-thought, and multi-agent orchestration as first-class capabilities [tweet 2].
Training showed predictable scaling across pretraining, RL, and test-time reasoning [tweet 2].
Contemplating mode orchestrates multiple agents reasoning in parallel, targeted at complex scientific and reasoning queries; internal testing reports it competitive with Gemini Deep Think and GPT Pro [tweet 3].
Pre-deployment safety evaluations cover frontier risk categories, behavioral alignment, and adversarial robustness, with reported strong refusal on bio/chem-weapon prompts [tweet 4].
Meta AI is now split into instant vs. thinking modes, plus a new shopping mode that uses creator/brand/styling signals across Meta apps [tweet 5].
Distribution surface is the existing ~3B users of Meta’s apps; private API preview is open today to select partners, with future versions planned to be open-sourced [tweet 6, tweet 7].

Method

No technical details are disclosed in the announcement thread — no parameter count, architecture, training data, training compute, or benchmark numbers. The thread is a product/positioning announcement, not a model card or tech report. The only operational details given are (a) the model is natively multimodal rather than retrofitted, (b) contemplating mode is an explicit multi-agent parallel-reasoning mode (separate from a default “thinking” mode), and (c) deployment is gated through Meta AI in Meta’s existing app surfaces plus a private API preview.

Results

No benchmark numbers are reported in the thread. The only quantitative-style claim is the qualitative “competitive with Gemini Deep Think & GPT Pro” framing for contemplating mode, sourced to Meta’s own internal testing [tweet 3]. The team author acknowledges rough edges to be polished over time [tweet 9].

Why it’s interesting

The release shape matters more than the (undisclosed) numbers: Muse Spark is positioned in the same closed-but-API-accessible cohort that Open foundation-model releases tracks as the comparison baseline against open releases like HunyuanImage 3.0 Technical Report and the Qwen3-VL family — Meta is now explicitly not shipping the first MSL model open, with open-source as a future-tense commitment. Contemplating mode is another datapoint in the Inference-Time Scaling “learned parallel-agent orchestration” axis previously populated by Kimi K2.5 Agent Swarm and GPT Pro / Gemini Deep Think — a fourth-major-lab entry in the same product slot, again without disclosed orchestrator details. The “predictable scaling across pretraining, RL, and test-time reasoning” framing also contrasts with On the Slow Death of Scaling, which argues that scaling is yielding less than headline framings suggest.