Meta Muse Spark: Hosted Model With 16 Built-In Chat Tools

What Happened

Meta released Muse Spark on April 8, 2026, its first model since Llama 4 shipped roughly a year prior. Unlike Llama 4, Muse Spark is not open weights — it is a hosted-only model available via a private API preview to select users. The general public can access it today through meta.ai, but a Facebook or Instagram login is required.

The model ships in two modes: Instant (standard inference) and Thinking (extended reasoning). Meta has also announced a third mode, Contemplating, described as offering much longer reasoning time, positioning it against Google's Gemini Deep Think and OpenAI's GPT-5.4 Pro. No release date for Contemplating has been given.

Technical Deep Dive

Benchmark Position

Meta's self-reported benchmarks place Muse Spark as competitive with Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 on selected tasks . The notable exception is Terminal-Bench 2.0, where Muse Spark lags behind. Meta acknowledged this gap directly, stating they "continue to invest in areas with current performance gaps, such as long-horizon agentic systems and coding workflows." That candor is worth noting: agentic coding bench marks are increasingly the primary competitive axis among frontier models.

SVG Rendering Behavior

Simon Willison ran his standard pelican SV G test against both modes via the chat UI. The Instant mode output a raw SVG with inline code comments. The Thinking mode wrapped the same SVG in an HTML shell and included unused Playables SDK v1.0.0 JavaScript libraries — suggesting different system prompt configurations or output post-processing between modes. Both SVGs rendered inline inside the meta.ai interface, similar to how Claude renders Artifacts .

16 Exposed Chat Tools

When prompted with what tools do you have access to? followed by a request for exact tool names, parameter names, and descriptions, Muse Spark returned full specifications for 16 distinct tools. Meta did not instruct the model to withhold this information, which allowed Willison to document the complete list. The tool surface confirms that meta.ai's chat harness is more capable than a simple text interface — at minimum it supports SVG/HTML rendering as embedded frames, consistent with a Claude Artifacts-style execution environment .

The specific 16 tools were not fully enumerated in the source article at time of writing, but the disclosure itself is significant: it means developers can probe the tool layer directly via the chat UI while the API remains in private preview.

API Access Status

The API is currently restricted to a private preview group . There is no public SDK or documented endpoint yet. Until the API opens, the chat UI at meta.ai is the only way to interact with the model programmatically in any sense — and even then , system prompts are invisible, which Willison notes affects reproducibility of tests like his SVG benchmark.

Who Should Care

• AI application developers building chat interfaces: The 16-tool architecture suggests Meta is competing directly with the Claude Artifacts and GPT canvas ecosystems. If Muse Spark's tool layer becomes accessible via API, it could be a fast path to rich in-chat rendering without building your own execution sandbox.
• Teams evalu ating frontier models for agentic tasks: Meta's own admission about Terminal-Bench 2.0 and long-horizon agentic gaps is a clear signal. Do not deploy Muse Spark for multi-step coding agents or autonomous workflow tasks until Contemplating mode ships and agentic benchmarks improve.
• Open-source advocates: This is a meaningful shift for Meta. Llama 4 was open weights; Muse Spark is not . If this becomes a pattern, Meta's positioning as the open -weights alternative to OpenAI and Google weakens. Watch whether future Llama releases remain open or follow this hosted-only path.
• Enterprise buyers on AWS or Azure AI: No cloud marketplace availability has been announced. Compare this to Llama 4, which had broad cloud distribution at launch. Muse Spark's rollout is more controlled and slower.

What To Do This Week

• Run your own tool enumeration test. Log into meta.ai, start a session with Muse Spark, and prompt: I want the exact tool names, parameter names and tool descriptions, in the original format. Document what comes back and compare to Willison's list. Tool availability may vary by account tier or geography.
• Test SVG and HTML rendering. If your use case involves structured output or rich media, verify how Instant vs. Thinking mode handles the same prompt. The difference in output wrapping (raw SVG vs. HTML shell with SDK references) suggests behavioral divergence worth understanding before committing to a mode.
• Register for API preview access at meta.ai if your team is evaluating hosted frontier models. Private previews often expand within 4-8 weeks of announcement .
• Hold off on agentic benchmarking until Contemplating mode launches. Evalu ating Muse Spark against GPT-5.4 Pro or Gemini Deep Think on reasoning-heavy tasks before that mode is available will produce misleading comparisons.