Meta launches Muse Spark, its first frontier model and first closed-weight AI system
Meta Superintelligence Labs has launched Muse Spark, a native multimodal reasoning model that scores 52 on the Artificial Analysis Intelligence Index, placing it in the top 5 frontier models. This marks Meta's first frontier-class model and its first AI system without open weights, representing a strategic shift from its open-source Llama strategy. The model achieves comparable efficiency to Gemini 3.1 Pro while matching Llama 4 Maverick capabilities with over an order of magnitude less compute.
How Muse Spark ranks now
| Benchmark | Score | Rank | Source |
|---|---|---|---|
| Arena Elo | 1489 Elo | 18 of 146 | LMArena, Nov 2026 ↗ |
| GPQA | 88.4% | 62 of 165 | Artificial Analysis, Jul 2026 ↗ |
| SWE-bench Verified | 74.4% | 42 of 80 | Vals AI, Sep 2026 ↗ |
Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.
Meta launches Muse Spark, its first frontier model and first closed-weight system
Meta Superintelligence Labs has unveiled Muse Spark, a native multimodal reasoning model that marks two significant departures from the company's AI strategy: it's Meta's first frontier-class model and its first system without open weights.
Benchmark Performance
Muse Spark scored 52 on the Artificial Analysis Intelligence Index, landing in the top 5 across all tested models. Only Gemini 3.1 Pro Preview (top performer), GPT-5.4, and Claude Opus 4.6 scored higher. For context, Meta's previous models Llama 4 Maverick and Scout achieved only 18 and 13 points respectively when they launched in April 2025.
Independent testing by Artificial Analysis shows the model closing the frontier gap in a single release. However, Artificial Analysis flagged weakness in agent-based tasks: on the GDPval-AA work task benchmark, Muse Spark scored 1,427 points versus Claude Sonnet 4.6's 1,648 and GPT-5.4's 1,676.
On Meta's internal testing, Muse Spark achieved 58% on Humanity's Last Exam and 38% on FrontierScience Research. In extended thinking mode without tools, it scored 50.2 on Humanity's Last Exam (No Tools), outperforming both Gemini 3.1 and GPT-5.4 Pro in this specific benchmark.
Key Capabilities and Architecture
Muse Spark operates as a native multimodal model with three core capabilities: tool usage, visual chain-of-thought reasoning, and multi-agent orchestration. The model includes a "Contemplating Mode" designed to compete with deep reasoning features in competing frontier models like Gemini Deep Think and GPT Pro.
Meta rebuilt the pretraining stack from the ground up over nine months, implementing changes to model architecture, optimization, and data curation. According to Meta's claims, Muse Spark matches Llama 4 Maverick's capabilities using over an order of magnitude less compute, positioning it as substantially more efficient than competing base models.
The company employs two approaches to test-time compute. The first uses thought-time penalties that optimize token consumption. Meta observed a phenomenon it calls "thought compression," where the model initially improves by thinking longer, then compresses reasoning to solve problems with fewer tokens before expanding solutions again for stronger results. The second approach uses multi-agent orchestration—deploying multiple parallel agents on difficult problems simultaneously—to boost performance without adding latency.
Artificial Analysis verified efficiency claims: Muse Spark consumed 58 million output tokens for the full Intelligence Index run, matching Gemini 3.1 Pro Preview (57 million) and well below Claude Opus 4.6 (157 million) or GPT-5.4 (120 million).
Closed Weights Mark Strategic Shift
Unlike the Llama family, Muse Spark is not open-weight and cannot run locally. This represents a sharp break from Meta's open-source playbook championed for years. Meta's AI chief Alexandr Wang stated the company has "plans to open-source future versions," suggesting closed weights may not be permanent policy. The company is also reportedly planning to open-source parts of its new AI models.
Meta justified the shift by noting its enormous spending on AI infrastructure and specialized talent "has to start paying for itself eventually."
Health and Multimodal Focus
Meta partnered with over 1,000 doctors to curate high-quality, factually accurate training data for health applications. The model can generate interactive displays breaking down nutritional value of food or showing which muscles activate during specific exercises. Meta emphasized multimodal perception and health as primary use cases, though interactive applications like mini-game generation are also possible.
Meta acknowledged performance gaps in long-horizon agentic systems and coding workflows. The company also flagged that Muse Spark frequently labeled test scenarios as "alignment traps" during security evaluation, demonstrating "evaluation awareness"—a phenomenon where models appear to recognize they're being tested.
Availability and Future Plans
Muse Spark is live on meta.ai and in the Meta AI app, with private API preview access going to select users. Pricing has not been disclosed.
Meta frames Muse Spark as "the first step on our scaling ladder and the first product of a ground-up overhaul of our AI efforts" toward "personal superintelligence." The company stated "bigger models are already in development with infrastructure scaling to match." This release follows a rough period for Meta's AI efforts after Llama 4 Maverick and Scout drew criticism in April 2025 for underwhelming benchmark results and internal accusations of benchmark manipulation.
What This Means
Muse Spark demonstrates Meta can compete at the frontier in a single leap, closing a gap that seemed substantial just months ago. However, persistent weaknesses in agentic tasks and the company's admission of gaps in coding workflows suggest the model may not be immediately ready for autonomous agent deployment. The shift to closed weights is pragmatic—Meta's infrastructure spending demands commercial revenue—but the stated commitment to open-sourcing future versions leaves the door open to returning to its original strategy. Real-world performance across extended reasoning tasks will be the critical test; benchmark scores alone may not reflect usability in production environments.
Related Articles
Thinking Machines Lab releases Inkling 2 Mini: 26B-parameter looped MoE with 3B active and 1M-token context
Thinking Machines Lab has released Inkling 2 Mini, a multimodal Mixture-of-Experts model with about 26B physical total parameters and 3B active per token. It repeats its middle layers three times, supports a 1M-token context, and ships with a DSpark drafter. Hugging Face reports it leads 32 of 48 benchmarks against four comparable models.
Tencent releases Hy4 flagship weights: 770B-parameter MoE, 1,048,576-token context, conditional license
Tencent released downloadable weights for its flagship Hy4 on November 20, with a System Card. The multimodal MoE has a 770-billion-parameter language backbone, 49 billion active parameters per text token and a 1,048,576-token context window, under a license with commercial conditions for model-as-a-service operators.
Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context
Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.
NVIDIA releases Nemotron-Labs-Sigma: 14 open 3B and 8B continuous diffusion language model checkpoints
NVIDIA released Nemotron-Labs-Sigma, a set of 14 open continuous diffusion language model checkpoints at 3B and 8B parameters. According to MarkTechPost, Sigma denoises text in a learned 16-D embedding space, distills to few-step samplers, and decodes up to 4.20x faster than autoregressive decoding.
Comments
Loading...