Meta launches Muse Spark model with private API preview and 16 integrated tools
Meta announced Muse Spark today, its first model release since Llama 4 a year ago. The hosted model is available in private API preview and on meta.ai with Instant and Thinking modes, benchmarking competitively against Anthropic's Opus 4.6 and Google's Gemini 3.1 Pro, though behind on Terminal-Bench 2.0.
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: First Major Model Since Llama 4
Meta announced Muse Spark today, marking its first significant model release since Llama 4 in April 2025. The model is available as a private API preview to select users and on meta.ai's chat interface, which requires Facebook or Instagram login.
Performance and Positioning
Meta's self-reported benchmarks show Muse Spark performing at parity with Anthropic's Opus 4.6, Google's Gemini 3.1 Pro, and OpenAI's GPT 5.4 on selected metrics. However, the model notably underperforms on Terminal-Bench 2.0. Meta explicitly acknowledges performance gaps in long-horizon agentic systems and coding workflows, stating they "continue to invest in areas with current performance gaps."
Deployment Modes
Muse Spark is accessible through two distinct modes on meta.ai:
Instant Mode: Faster inference with direct SVG/HTML output capabilities. Testing showed this mode generates SVG code directly with embedded comments.
Thinking Mode: Extended reasoning with wrapped output and enhanced problem-solving. Early benchmarking demonstrates notably better performance on complex tasks compared to Instant mode.
Meta plans a future Contemplating Mode offering significantly longer reasoning times, comparable to Google's Gemini Deep Think and OpenAI's GPT-5.4 Pro.
16 Integrated Tools Exposed
Meta's chat interface includes 16 publicly accessible tools. Notably, Meta did not restrict disclosure of these tools, allowing direct enumeration without workarounds.
Web Capabilities:
browser.search: Web search through undisclosed enginebrowser.open: Full page retrieval from search resultsbrowser.find: Pattern matching on page content
Meta Platform Integration:
meta_1p.content_search: Semantic search across Instagram, Threads, and Facebook posts (user-accessible content only, post-2025-01-01)meta_1p.meta_catalog_search: Product catalog searchcontainer.download_meta_1p_media: Pull media from Meta platforms into sandbox
Code Execution and Artifacts:
container.python_execution: Code Interpreter running Python 3.9.25 with pandas, numpy, matplotlib, plotly, scikit-learn, PyMuPDF, Pillow, and OpenCV. Note: Python 3.9 reached end-of-life status in October 2024.container.create_web_artifact: HTML/JavaScript sandboxed rendering (Claude Artifacts-style)container.visual_grounding: Image analysis with object detection, bounding box generation, point localization, and object counting
File Operations:
container.view,container.insert,container.str_replace: Text editor commands mirroring Claude's implementationcontainer.file_search: Semantic search across uploaded files
Media and Agents:
media.image_gen: Image generation with "artistic" and "realistic" modes; returns "square", "vertical", or "landscape" formatssubagents.spawn_agent: Sub-agent spawning for delegated research and analysisthird_party.link_third_party_account: Account linking for Google Calendar, Outlook Calendar, Gmail, Outlook
Technical Observations
Direct testing revealed Python 3.9.25 and SQLite 3.34.1 (January 2021 release) running in the sandbox environment. The Thinking mode's tendency to wrap SVG in HTML shells with Playables SDK v1.0.0 JavaScript libraries suggests integration with Meta's interactive content framework.
What This Means
Muse Spark positions Meta as a credible competitor in the hosted LLM space, though the private preview status limits immediate adoption. The 16-tool architecture demonstrates Meta's integration strategy: leveraging Facebook, Instagram, and Threads data access as competitive moat while matching feature parity with Claude and ChatGPT. The acknowledged gaps in agentic systems and coding—combined with Python 3.9 EOL tooling—suggest this release prioritizes chat functionality over agent-grade reliability. The planned Contemplating mode indicates Meta is responding directly to Deep Research and Pro reasoning mode adoption patterns.
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...