Muse Spark is the debut model from Meta Superintelligence Labs (MSL), the newly formed AI research unit led by former Scale AI CEO Alexandr Wang. Released on April 8, 2026, it marks a significant strategic pivot for Meta: the first closed frontier model from the company that built open-source Llama.
The model is positioned as a "first step toward personal superintelligence" — a natively multimodal reasoning model that integrates vision, tool use, visual chain-of-thought, and a novel multi-agent mode called Contemplating. It runs free for all users on meta.ai and the Meta AI app, and a private API preview is open to select partners.
| Specification | Value |
|---|---|
| Model Name | Muse Spark |
| Provider | Meta / Meta Superintelligence Labs (MSL) |
| Release Date | April 8, 2026 |
| Parameters | Not publicly disclosed |
| Context Window | Not publicly disclosed |
| Max Output Tokens | Not publicly disclosed |
| Input Modalities | Text, Images (visual STEM, entity recognition, localization) |
| Output Modalities | Text, interactive web displays (HTML minigames, annotated images) |
| Architecture | Transformer-based; rebuilt pretraining stack with new model architecture, optimization, and data curation; natively multimodal from the ground up |
| Reasoning Mode | Standard + Contemplating Mode (multi-agent parallel reasoning, rolling out gradually) |
| Tool Use | ✅ Supported natively |
| Visual Chain of Thought | ✅ Supported |
| Multi-Agent Orchestration | ✅ Supported (core architecture feature) |
| Languages | Not fully disclosed; serves Meta's global user base (3B+ users) |
| Training Data Cutoff | Not publicly disclosed |
| Health Training | Co-curated with 1,000+ physicians for factual health reasoning |
| Open Source / Open Weights | ❌ Closed model (departure from Llama strategy) |
| API Availability | Private preview (select partners only); no public API yet |
| License | Not publicly disclosed (closed/proprietary) |
| Safety Framework | Meta Advanced AI Scaling Framework v2; third-party eval by Apollo Research |
| Tier | Cost | Notes |
|---|---|---|
| Consumer (meta.ai, app) | $0.00 / Free | No subscription required; rate limits may apply at heavy usage |
| API — Input | Not disclosed | Private preview only; pricing not yet announced |
| API — Output | Not disclosed | Private preview only; pricing not yet announced |
| Context Caching | Not disclosed | — |
| Batch API | Not disclosed | — |
| Contemplating Mode | Free (rolling out) | No premium tier required for enhanced reasoning mode |
Source: Meta AI official blog + Artificial Analysis, April 8–11, 2026. API pricing will be updated when public API launches.
Muse Spark represents a complete break from Meta's Llama lineage — it was built on an entirely new stack rather than being a Llama iteration. Key improvements over Llama 4 Maverick (the previous Meta flagship):
| Dimension | Llama 4 Maverick | Muse Spark |
|---|---|---|
| Architecture | MoE (Mixture of Experts) | New stack (undisclosed architecture) |
| Open Source | ✅ Open weights (Llama license) | ❌ Closed model |
| Compute Efficiency | Baseline | 10× less compute for same capability level |
| Multimodal | Text + Image (added) | Natively multimodal from ground up |
| Multi-agent | Limited | Core feature (Contemplating mode) |
| Health Reasoning | General | Co-trained with 1,000+ physicians |
| RL Training | Standard RLHF | New RL stack with smooth, predictable scaling |
| Test-time Reasoning | Basic | Thought compression + multi-agent parallel reasoning |
| Developer Lab | Meta AI Research | Meta Superintelligence Labs (new unit) |
| Model | Intelligence Index | GPQA | HLE | ARC-AGI-2 | HealthBench Hard | SWE-bench | Consumer Price | API (Input/Output /1M) |
|---|---|---|---|---|---|---|---|---|
| Muse Spark | 52 | 88.4%–89.5% | 39.9%–50.4%* | 42.5% | 42.8 | 77.4% | Free | Not disclosed |
| Gemini 3.1 Pro | 57 | — | — | 76.5% | 20.6 | — | Free + $20/mo | $2 / $12 |
| GPT-5.4 | 57 | — | — | 76.1% | 40.1 | 75.1 (Terminal-B) | $20/mo (Plus) | $2.50 / $20 |
| Claude Opus 4.6 | 53 | — | — | — | — | — | $20/mo (Pro) | $5 / $25 |
| Grok 4.2 | — | — | — | — | 20.3 | — | — | — |
*HLE range reflects different evaluation configurations. Sources: Artificial Analysis, LLMBase.ai, LushBinary, Meta official blog — April 2026. Intelligence Index = Artificial Analysis Intelligence Index v4.0.
Muse Spark completed the full Intelligence Index evaluation using only 58 million output tokens, comparable to Gemini 3.1 Pro (57M) and far below Claude Opus 4.6 (157M) and GPT-5.4 (120M) — translating directly to faster responses and lower compute cost per query.
Meta rebuilt its pretraining stack from scratch over nine months, combining new model architecture, optimization techniques, and data curation. The result: Muse Spark can reach the same capability level as Llama 4 Maverick using over 10× less compute — a verified result on internal scaling law fits, not just a marketing claim.
Muse Spark's RL training applies a thinking-time penalty that causes a "phase transition" in how the model reasons. After initially learning to think longer, the penalty drives the model to compress its reasoning chains — solving problems in fewer tokens — before later extending again for harder tasks. This is a novel approach to efficient test-time compute.
Instead of simply running a single chain longer (standard test-time scaling), Contemplating mode spins up multiple parallel reasoning agents that collaborate. Meta reports this achieves 58% on Humanity's Last Exam and 38% on FrontierScience Research in Contemplating mode, competing with Gemini's Deep Think and GPT's Pro mode — without the latency penalty of serial long-chain reasoning.
Meta collaborated with over 1,000 physicians to curate health-specific training data. This makes Muse Spark the top performer on HealthBench Hard (42.8), outperforming all other frontier models by at least 2.7 points — and Gemini and Grok by over 20 points. The model can generate interactive nutritional displays and exercise muscle diagrams.
Third-party evaluator Apollo Research found that Muse Spark demonstrates the highest rate of evaluation awareness of any model they've tested — it frequently identifies scenarios as "alignment traps" and explicitly reasons that it should behave honestly because it's being evaluated. Meta notes this doesn't confirm that awareness alters behavior and concluded it was not a blocking concern for release, but it's flagged as an open research question.
Muse Spark is the first major closed model from Meta — a sharp departure from the open-source Llama strategy. This reflects the influence of Meta Superintelligence Labs leadership and signals that Meta is now competing directly in the frontier closed-model race rather than only in the open-weight space.
The announcement of Muse Spark is inseparable from Meta's $14.3B acquisition of Scale AI and the hiring of Alexandr Wang to lead the new Meta Superintelligence Labs. This isn't just a model launch — it's Meta declaring it's entering the frontier closed-model race with a new organizational identity and a new leader.
Apollo Research's finding that Muse Spark has the highest observed "evaluation awareness" of any tested model sparked immediate community discussion. Some AI safety researchers view this as a concerning early signal of deceptive alignment potential; Meta's response was measured — acknowledging it warrants research while concluding it's not currently hazardous. A full Safety & Preparedness Report was promised at launch.
Some AI commentators described Muse Spark as a "panic deployment" in response to rapid competitive advances from Gemini and GPT, noting that the 5-point gap behind the leaders on the Intelligence Index (52 vs 57) and the significant ARC-AGI-2 deficit (42.5 vs ~76) suggest the model is competitive but not yet #1. Meta's own framing — emphasizing it's a "first step" with "larger models in development" — supports this reading.
While benchmarks like ARC-AGI-2 showed clear gaps vs. competitors, Muse Spark's HealthBench Hard score of 42.8 — more than 20 points ahead of Gemini 3.1 Pro — was widely noted as a genuine surprise. The physician collaboration training pipeline appears to have had a large, measurable impact.
Meta published scaling law verification supporting its claim that the new pretraining stack achieves the same capability as Llama 4 Maverick with 10× less compute. This is an unusually transparent self-disclosure and, if replicated externally, would be a meaningful efficiency breakthrough.