概要
Muse Spark is the debut model from Meta Superintelligence Labs, representing a strategic pivot from Meta's open-source Llama lineage to a proprietary, natively multimodal reasoning architecture. Launched in April 2026, the model was built from the ground up over nine months following Meta's $14.3 billion investment in Scale AI and the establishment of MSL under Alexandr Wang. It scores 51 on the Artificial Analysis Intelligence Index, placing it among the frontier tier but behind GPT-5.4, Gemini 3.1 Pro, and Claude Opus 4.6 on general benchmarks.
The model's standout innovations include its multi-agent 'Contemplating' mode—which orchestrates parallel reasoning agents to achieve superior performance on complex tasks like Humanity's Last Exam (62.1% with tools)—and exceptional performance in health and medical AI, where it leads all competitors with a 42.8% score on HealthBench Hard. Architecturally, Muse Spark is natively multimodal from the ground up, integrating text, image, and audio processing rather than bolting vision onto a language backbone. This enables unique capabilities like visual chain-of-thought reasoning and interactive health visualizations.
While Muse Spark excels in health, scientific reasoning, and multimodal perception, it currently lags significantly in coding, abstract reasoning, and agentic computer-use tasks compared to GPT-5.4 and Claude. The model is available for free to consumers through Meta's apps, with a private API preview for developers. Meta positions it as the first step on a scaling ladder toward 'personal superintelligence,' with larger models already in development. Its token efficiency—using roughly half the output tokens of competitors for comparable benchmark runs—makes it economically viable for Meta to offer at no cost across its 3+ billion user base.
ベンチマーク＆性能
Muse Spark 1.1 demonstrates strong performance in specific domains while showing gaps in others compared to frontier models. The table below summarizes key benchmark scores from Meta's official documentation and third-party evaluations.
| Benchmark | Muse Spark 1.1 (Meta) | GPT-5.4 (OpenAI) | Claude Opus 4.6 (Anthropic) | Gemini 3.1 Pro (Google) | Notes |
|-----------|----------------------|------------------|----------------------------|------------------------|-------|
| **Intelligence Index (AA v4.0)** | 51 | 57 | 53 | 57 | Muse Spark 1.1 tied with GLM-5.2 and GPT-5.6 Luna |
| **Humanity's Last Exam (with tools)** | 62.1% | 52.2% | 57.9% | 51.4% | Muse Spark 1.1 leads with tools |
| **HealthBench Hard** | 42.8% | 40.1% | 14.8% | 20.6% | Muse Spark dominates health reasoning |
| **CharXiv Reasoning (Charts)** | 88.4% | 82.8% | 65.3% | 80.2% | Strong visual understanding |
| **GPQA Diamond (PhD-level)** | 89.5% | 92.8% | 92.7% | 94.3% | Competitive but not leading |
| **ARC-AGI-2 (Abstract Reasoning)** | 42.5% | 76.1% | 63.3% | 76.5% | Significant weakness |
| **SWE-Bench Verified (Coding)** | 77.4% | - | 80.8% | 80.6% | Behind leaders |
| **Terminal-Bench 2.0 (Agentic Coding)** | 59.0% | 75.1% | 65.4% | 68.5% | Major gap |
| **GDPval-AA Elo (Office Tasks)** | 1444 | 1672 | 1606 | 1320 | Weakest in agentic tasks |
| **DeepSWE 1.1 (Long-horizon Coding)** | 53.3% | 67.0% | 59.0% | 12.0% | Context-dependent |
Key observations:
1. **Health & Multimodal Leadership**: Muse Spark leads by wide margins on HealthBench Hard and CharXiv Reasoning, reflecting its specialized training and native multimodality.
2. **Contemplating Mode Advantage**: On Humanity's Last Exam with tools, Muse Spark 1.1's 62.1% score leads GPT-5.4's 52.2%, demonstrating the power of multi-agent parallel reasoning.
3. **Coding & Agentic Gaps**: Terminal-Bench 2.0 and ARC-AGI-2 scores show significant weaknesses in agentic coding and abstract reasoning compared to GPT-5.4 and Claude.
4. **Token Efficiency**: Muse Spark completed the full Intelligence Index evaluation using 94M output tokens versus 109M for GPT-5.4 and 141M for GLM-5.2, indicating superior computational efficiency.
*Sources: Meta AI blog, Artificial Analysis Intelligence Index v4.0, jdhodges.com benchmark analysis*
詳細比較
**Muse Spark vs GPT-5.4**
- **Pricing**: Muse Spark is free for consumers; GPT-5.4 requires $20/month subscription. API: Muse Spark $1.25/$4.25 vs GPT-5.4 $2.50/$20.
- **Context Window**: Both offer 1M tokens.
- **Strengths**: Muse Spark leads in health (42.8% vs 40.1%), chart understanding (88.4% vs 82.8%), and token efficiency. GPT-5.4 dominates in coding (75.1 vs 59.0 Terminal-Bench), abstract reasoning (76.1 vs 42.5 ARC-AGI-2), and agentic tasks (1672 vs 1444 GDPval-AA Elo).
- **When to choose**: Muse Spark for health applications, cost-sensitive deployments, and Meta ecosystem integration. GPT-5.4 for coding, autonomous agents, and abstract reasoning.
**Muse Spark vs Claude Opus 4.6**
- **Pricing**: Muse Spark free vs Claude $20/month. API: $1.25/$4.25 vs $5/$25.
- **Context Window**: Both 1M tokens.
- **Strengths**: Muse Spark leads in health (42.8% vs 14.8%) and multimodal reasoning. Claude leads decisively in coding (80.8% vs 77.4% SWE-Bench Verified) and has more mature agentic features with Agent Teams.
- **When to choose**: Muse Spark for health, scientific research, and consumer applications. Claude for software engineering, code review, and production agentic workflows.
**Muse Spark vs Gemini 3.1 Pro**
- **Pricing**: Muse Spark free vs Gemini free tier + $20/month. API: $1.25/$4.25 vs $2/$12.
- **Context Window**: Both 1M tokens.
- **Strengths**: Muse Spark leads in health and chart understanding. Gemini leads in overall intelligence (57 vs 51), abstract reasoning (76.5 vs 42.5 ARC-AGI-2), and offers better API value.
- **When to choose**: Muse Spark for health, multimodal perception, and free consumer access. Gemini for cost-effective API usage, general reasoning, and scientific tasks.
**Key Differentiator**: Muse Spark's unique 'Contemplating' mode provides an architectural advantage for complex reasoning through parallel agents, though it's still rolling out. Its native multimodality and health specialization create niche advantages despite general benchmark gaps.
コミュニティ評価
Developer and researcher reactions highlight both excitement about Meta's new direction and pragmatic concerns about current limitations:
**Positive Reception**:
- **Enterprise Developers**: Early partners praise Muse Spark as a 'complete agentic foundation' (Replit CEO Amjad Masad) with strong tool use and long context handling.
- **AI Researchers**: The 10x compute efficiency claim and novel multi-agent Contemplating mode have generated significant interest in the research community.
- **Health Tech**: The physician-curated training data and HealthBench Hard leadership are seen as a strategic moat in an emerging $45B+ market.
**Critical Concerns**:
- **Open-Source Community**: The shift from open Llama to proprietary Muse Spark has disappointed developers who built stacks assuming Meta's frontier models would remain open-weight.
- **Production Developers**: Limited API access (private preview only) and acknowledged coding/agentic gaps mean most developers can't build with it yet.
- **Benchmark Skeptics**: Some analysis notes that Muse Spark only wins 3 of 20 benchmarks in Meta's own comparison chart, with Gemini leading 8 and GPT-5.4 leading 6.
**Adoption Patterns**:
- **Consumer Adoption**: Rapid integration across Meta's 3+ billion user base via Facebook, Instagram, WhatsApp, and Meta AI app.
- **Enterprise Interest**: Early adoption in health tech, e-commerce (Facebook Marketplace automation), and content creation workflows.
- **Developer Hesitation**: Most are waiting for public API availability and evidence of closing coding/agentic gaps before production deployment.
The consensus is that Muse Spark represents a strategically important but currently incomplete offering, with its true potential dependent on Meta's ability to scale the architecture while closing performance gaps.
ユースケース
1. **Health & Wellness Applications**: Muse Spark's physician-curated training enables interactive health visualizations—analyzing food photos for nutritional breakdown, explaining muscle groups during exercise, or personalizing health recommendations. For example, a user can photograph their fridge contents and receive a week-long meal plan aligned with specific dietary goals like managing cholesterol.
2. **Scientific Research & Complex Reasoning**: The Contemplating mode excels at multi-agent parallel reasoning for problems like literature review, experimental design, or multidisciplinary analysis. Researchers can use it to synthesize findings across papers, identify contradictions, and generate hypotheses, as demonstrated by its 38.3% score on FrontierScience Research.
3. **Multimodal Content Creation & Analysis**: Native vision capabilities enable interactive troubleshooting (pointing a camera at a broken appliance to get annotated repair guides), chart/data visualization interpretation, and visual-to-code generation. Developers can build applications where users describe interfaces verbally while showing screenshots, with the model generating corresponding code.
4. **Large-Scale Consumer AI Integration**: As a free model powering Meta's entire ecosystem, it's ideal for adding intelligent features to social platforms—analyzing marketplace listings from smartphone videos, generating interactive tutorials from product images, or providing real-time visual search across billions of photos and posts.
**When to Choose Muse Spark Over Alternatives**:
- Choose Muse Spark when health accuracy, cost efficiency, or Meta ecosystem integration are priorities.
- Choose GPT-5.4 when coding quality, abstract reasoning, or autonomous agentic tasks are critical.
- Choose Claude Opus 4.6 for software engineering workflows or when mature agent orchestration is needed.
- Choose Gemini 3.1 Pro for cost-effective API usage on general reasoning tasks.
最新ニュース
1. **Muse Spark 1.1 Release (July 9, 2026)**: Major update with 8-point gain on Intelligence Index, 1M context window, and public preview of Meta Model API. Significant improvements in coding, agentic tasks, and computer use.
2. **Public API Preview Launch**: Developers can now access Muse Spark 1.1 through the new Meta Model API (OpenAI-compatible), though still in public preview for US developers.
3. **Pricing Announcement**: API pricing set at $1.25/$4.25 per 1M input/output tokens, with cache hits discounted to $0.15/1M. Consumer access remains free.
4. **Contemplating Mode Rollout**: The multi-agent parallel reasoning mode is gradually rolling out in Meta AI, achieving 58% on Humanity's Last Exam with tools.
5. **Ecosystem Integrations**: Muse Spark is now integrated across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta AI glasses, reaching 3+ billion users.
6. **Safety & Preparedness Report Release**: Meta published detailed safety evaluations following the Advanced AI Scaling Framework, showing Muse Spark operates within safe margins across frontier risk categories.
7. **Enterprise Partnerships**: Early access partnerships with companies like Replit, Cline, and Box, who report competitive enterprise capabilities for structured workflows.
8. **Scaling Roadmap Announcement**: Meta confirmed Muse Spark is the first step in a scaling program, with larger models already in development and supported by Hyperion data center investments.
9. **Research Investments**: Meta continues investing in closing performance gaps, particularly in long-horizon agentic systems and coding workflows where Muse Spark currently trails competitors.