Developer
Meta

Muse Spark

Meta's frontier model, released April 8, 2026 by Meta Superintelligence Labs. A natively multimodal reasoning model with tool use and multi-agent orchestration.

Released
Apr 8, 2026
Type
Multimodal Reasoning Model
License
Proprietary
🌐 Official Website
On this page

Overview

Muse Spark is a Meta model. It was built by Meta Superintelligence Labs (MSL) and released on April 8, 2026 — announced the same day on the AI at Meta blog under the title "Introducing Muse Spark: Scaling Towards Personal Superintelligence" and on the Meta Newsroom as "Meta's Most Powerful Model Yet."

Meta describes it as "the first in a new series of large language models built by Meta Superintelligence Labs," and technically as a natively multimodal reasoning model with support for tool use, visual chain of thought, and multi-agent orchestration.

The strategic significance is larger than the model. Muse Spark is closed-weights — a deliberate break from the open-weight lineage that defined Llama. Meta has said it "hope[s] to open-source future versions of the model," language that concedes this one is not open. For a company whose AI identity was built on releasing downloadable weights, shipping a proprietary frontier flagship is the story.

Muse Spark supersedes Llama 4 as Meta's frontier model. Llama continues as Meta's open-weight family; Muse is now where Meta puts its best.

Capabilities

  • Native multimodality — Meta describes Muse Spark as a "natively multimodal reasoning model," not a language model with vision bolted on. Meta cites complex reasoning across science, math, and health domains, with strong multimodal perception.
  • Visual chain of thought — the model reasons through visual input step by step. See chain-of-thought reasoning for the underlying technique.
  • Tool use — first-class support, rather than a wrapper behaviour.
  • Multi-agent orchestration — the architectural feature that Contemplating mode exposes to users.
  • Contemplating mode — Meta's reasoning mode, which "orchestrates multiple agents that reason in parallel." Meta rolled it out gradually to meta.ai after launch. Muse Spark's headline benchmark results are Contemplating-mode results.
  • Pretraining efficiency — Meta claims it can "reach the same capabilities with over an order of magnitude less compute" compared with Llama 4 Maverick, and that Muse Spark is "significantly more efficient than the leading base models."

That last point is the most consequential technical claim Meta makes. A greater-than-10x reduction in the compute needed to reach a given capability level is a statement about Meta's pretraining recipe, not about Muse Spark's ceiling.

Technical Specifications

Meta has published very little conventional specification for Muse Spark. We list only what Meta has stated.

  • Developer: Meta Superintelligence Labs (MSL), part of Meta
  • Model type: natively multimodal reasoning model
  • Reasoning: Contemplating mode, orchestrating multiple agents reasoning in parallel
  • Supported: tool use, visual chain of thought, multi-agent orchestration
  • Weights: closed. No download, no self-hosting.
  • Architecture: Not disclosed. Meta has not published a parameter count, layer configuration, or whether the model uses a mixture-of-experts design.
  • Context window: Not published by Meta. A 262,144-token figure appears on aggregator sites with no Meta source behind it; we do not report it.
  • Knowledge cutoff: Not published by Meta.
  • Pricing: Not disclosed.

The absence of a published context window is genuine, not an oversight on our part. Meta's launch materials are written for a consumer and partner audience, and the company has not released a model card with the specifications a developer would expect.

Use Cases

  • Complex reasoning in science, math, and health — the three domains Meta names explicitly.
  • Research-grade question answering — the FrontierScience Research and Humanity's Last Exam results position the model against hard, expert-level questions.
  • Visual reasoning tasks — problems where the reasoning must run through an image, not merely start from one.
  • Tool-augmented workflows — where the model must call out to external capabilities mid-reasoning.
  • Parallel multi-agent problem solving — Contemplating mode's native shape: decompose, reason in parallel, reconcile.
  • Consumer assistant — the model is free at meta.ai and in the Meta AI app, and rolls out across Meta's messaging surfaces and AI glasses.

Performance / Benchmarks

All figures are Meta-reported, from the Muse Spark launch materials, and all are Contemplating-mode results.

BenchmarkMuse Spark (Contemplating mode)
Humanity's Last Exam58%
FrontierScience Research38%

Compute efficiency: Meta claims it can "reach the same capabilities with over an order of magnitude less compute" than Llama 4 Maverick, and that Muse Spark is "significantly more efficient than the leading base models."

Meta has not published results on standard developer benchmarks such as SWE-bench, MMLU, or GPQA, and has not released a comparison table against competing frontier models. Independent evaluation is limited by the fact that API access remains in private preview.

Limitations

  • No open weights. For a Meta model this is the defining constraint. No self-hosting, no fine-tuning, no offline deployment, no inspection. Meta hopes to open-source future versions; it has not committed to it.
  • No API for most developers. Access is a private preview limited to select partners. There is no self-serve route and no published pricing.
  • No published specifications. No context window, no parameter count, no knowledge cutoff, no architecture disclosure, no model card.
  • Narrow benchmark disclosure. Two benchmark numbers, both in Contemplating mode. Nothing on coding, nothing on standard academic suites, no head-to-head comparison table.
  • Little independent verification. Because API access is gated, third-party evaluation of Meta's claims is scarce.
  • Contemplating mode rolled out gradually. The mode behind the headline numbers was not available to all meta.ai users at launch.
  • Consumer-first framing. The launch materials address consumers and partners, not developers. Anyone planning an integration is working with incomplete information.

Pricing & Access

Meta has not disclosed pricing for Muse Spark.

  • Free consumer access — available now at meta.ai and in the Meta AI app.
  • API — private preview only. At launch Meta stated: "It will be available in private preview via API to select partners." No public terms, no self-serve signup, no published rate card.
  • Meta's apps — rolling out to WhatsApp, Instagram, Facebook, and Messenger.
  • AI glasses — rolling out to Ray-Ban Meta and Oakley Meta AI glasses, with Meta Ray-Ban Display slated for summer 2026.

Ecosystem & Tools

Muse Spark anchors an expanding Muse line. On July 7, 2026 — three months after Muse Spark — Meta launched Muse Image and previewed Muse Video, described as the first media generation models developed by Meta Superintelligence Labs.

  • Muse Image — Meta's image generation model. Rather than mapping prompts directly to images, it operates as an agent: it invokes search and coding tools to improve accuracy, self-refines its generations, and improves through scaling test-time compute. It integrates with Muse Spark, draws on Instagram for social context, and embeds Content Seal, an invisible watermarking system for verifying AI-generated images. Meta reports it holds the No. 2 spot on Arena for text-to-image, single-image editing, and multi-image editing by human-preference Elo. Available in the Meta AI app, on meta.ai, in Instagram Stories in the US, and on WhatsApp in limited countries.
  • Muse Video — built on the same pretraining base as Muse Image, with native audio support. Meta reports No. 3 in human-preference Elo for text-to-video. Coming soon to creators and Meta AI.
  • Llama 4 — Meta's open-weight family. Muse Spark supersedes it as the frontier flagship, while Llama continues to serve open deployment.
  • meta.ai — the free consumer surface for Muse Spark.
  • Meta AI app — the mobile client, and where Contemplating mode rolled out.

Community & Resources

Frequently Asked Questions

Meta. It was built by Meta Superintelligence Labs (MSL), Meta's frontier research organisation, and Meta describes it as "the first in a new series of large language models built by Meta Superintelligence Labs" and as "Meta's most powerful model yet." It is not a ByteDance, OpenAI, or Google model.
April 8, 2026. Meta announced it simultaneously on the AI at Meta blog and the Meta Newsroom.
No. Muse Spark is a new, closed-weights architecture and a deliberate break from Meta's open-weight Llama lineage. It is the flagship of a new model series, not a Llama release. Meta has said it "hope[s] to open-source future versions of the model," which implies Muse Spark itself is not open.
No. Muse Spark is closed-weights. This is the significant strategic change from Llama, which Meta released with downloadable weights. Meta has stated it hopes to open-source future versions, but has made no commitment for Muse Spark.
A reasoning mode that, in Meta's words, "orchestrates multiple agents that reason in parallel." It is where Muse Spark's strongest benchmark results come from, and Meta rolled it out gradually to meta.ai after launch.
In Contemplating mode, Meta reports 58% on Humanity's Last Exam and 38% on FrontierScience Research. Meta also claims it can "reach the same capabilities with over an order of magnitude less compute" compared with Llama 4 Maverick.
Meta has not published one. A 262,144-token figure circulates on aggregator sites but does not appear in any Meta statement, so we do not report it.
Meta has not disclosed pricing. Muse Spark is free to use at meta.ai and in the Meta AI app. API access is in private preview to select partners, with terms not made public.
Only through Meta's private preview. At launch Meta said "it will be available in private preview via API to select partners." There is no self-serve API and no public pricing.
Meta's first media generation models from Meta Superintelligence Labs, launched July 7, 2026. Muse Image integrates with Muse Spark and uses agentic tool use — it invokes search and coding tools, self-refines its generations, and scales test-time compute. Muse Video is built on the same pretraining base and adds native audio. They mark the Muse line expanding beyond language.
As Meta's frontier flagship, yes — Muse Spark supersedes Llama 4 at the top of Meta's lineup. Llama remains Meta's open-weight family, and Meta has not announced its discontinuation. The two now serve different strategic roles: Llama for open deployment, Muse for frontier capability.

Related Models

Llama 4

Meta's last open-weight frontier release: Scout and Maverick, natively multimodal Mixture-of-Experts models with 17B active and up to a 10M token context.

GPT-5.6

OpenAI's flagship family — Sol, Terra and Luna — generally available since July 9, 2026. All three share a 1,050,000 token context and a February 2026 cutoff.

Claude Opus 4.8

Anthropic's most capable Opus-tier model — highly autonomous, state-of-the-art on long-horizon agentic work, knowledge work, and memory, with 1M token context.

Gemini 3.5

Google's Gemini 3.5 generation. Gemini 3.6 Flash, shipped July 21, 2026, is the flagship: 1M token context, a March 2026 cutoff, $1.50/$7.50 per 1M.

Grok 4.5

xAI's flagship model, a mixture-of-experts system trained jointly with Cursor. A 500K token context at $2.00 in / $6.00 out per million tokens.

Explore More Models

Discover other AI models and compare their capabilities.

Browse Models← Back to Models