Muse Spark

Learning Objectives

• Understand what Muse Spark is and how it fits into Meta's AI strategy
• Explain the significance of Meta Superintelligence Labs and the hiring of Alexandr Wang
• Evaluate Muse Spark's capabilities and its relationship to the Llama model family

What Is Muse Spark?

Muse Spark is Meta's flagship AI model, announced on April 8, 2026. It is the first product from Meta Superintelligence Labs — Meta's new internal AI research organization — and represents Meta's most capable model to date.

Code-named "Avocado" during development, Muse Spark was built over approximately 9 months. The model accepts voice, text, and image inputs (text output only at launch) and features multiple operating modes including a fast mode, several reasoning modes, and a shopping mode designed for commerce applications across Meta's platforms.

Muse Spark 1.1 and the Meta Model API (July 2026)

On July 9, 2026, Meta released Muse Spark 1.1 and used it to enter the crowded AI coding market — a direct challenge to Cursor, Claude Code, and OpenAI's GPT-5.6. The 1.1 update reframes Muse Spark as a multimodal reasoning model built for agentic tasks, with three headline strengths:

• Coding — Meta says the model improved substantially on enterprise codebases, able to diagnose and fix complex bugs, implement features in enterprise-grade systems, and execute large code migrations
• Agents and automation — it manages complex projects by delegating work across parallel subagents, actively managing a one-million-token context window to keep long tasks on track
• Computer use — it "understands when to automate and when to use the interface directly," writing scripts for efficiency and falling back to direct interaction when that is simpler

The release also brings multimodal input (images, video, PDFs), structured output, parallel tool calling, zero-shot generalization to new tools and custom skills, and built-in search with citations.

The distribution change is the strategic part: Muse Spark 1.1 launched in public preview through the new Meta Model API, which ships with an OpenAI-compatible package so developers can point existing code at it with minimal changes. That moves Meta from a consumer-AI provider toward a developer platform competing head-on for agentic-coding workloads.

Muse Spark 1.2 and Muse Code (August 2026)

On August 5, 2026, less than a month after 1.1, Meta shipped Muse Spark 1.2 together with Muse Code, a terminal coding agent the model was co-trained to work inside. That pairing is the release's central claim: the published scores describe the model and the agent together, not the model in isolation.

Meta says 1.2 was trained on substantially more coding compute with broader task diversity, focusing on whole-repository generation and large end-to-end projects. It also used Muse Spark 1.1 to generate harder training environments for its successor — a self-improvement loop applied to building training data, not to the model's weights at run time.

Pricing through the Meta Model API splits into two tiers, and the split is the notable part. Standard runs $1.25 per million input tokens and $4.25 per million output. Contributor drops to 10 cents and 20 cents respectively — roughly 12-times cheaper on input — in exchange for Meta using the code you send to improve its models. For a personal project that may be a fair trade; under a client contract or an employer's intellectual-property policy it usually is not a developer's decision to make alone.

Muse Spark 1.3 (September 2026)

On September 2, 2026, less than a month after 1.2, Meta shipped Muse Spark 1.3 on the Meta Model API and the Muse Code command line. The stated focus is agentic reliability rather than raw capability: Meta says the model asks clarifying questions when a prompt is ambiguous, calls for help when it is stuck, and confirms before taking consequential actions, and that it is more realistic about what it can and cannot do rather than burning tokens down dead-end paths.

That framing is worth reading carefully, because it describes a model designed to do less on its own initiative. The benefit Meta claims is economic as much as behavioural — fewer turns and fewer tokens to finish a task, which makes an already inexpensive frontier model cheaper still.

Chief AI Officer Alexandr Wang positions 1.3 as competitive with Claude Fable 5.1 and better than GPT-5.6 Sol at coding. Independent benchmarking by Artificial Analysis puts 1.3 roughly level with GPT-5.6 Sol, Claude Opus 5 and Grok 4.6 High on overall intelligence — a four-point jump over 1.2 — which is the rare case of third-party numbers broadly supporting a vendor's own claim. Treat "better at coding" as the unverified half.

The open-weights question is still open

On the same day, Mark Zuckerberg promised open weights for Muse Spark "soon" in a post on X. That promise names no version, and the detail matters: Meta has not decided whether to publish 1.3's weights, and still plans to release only those for 1.2. So the honest statement of where things stand is that Meta has committed to the idea of an open Spark without committing to a specific checkpoint or date.

There is a regulatory dimension worth knowing, because it limits how much the label buys Meta. Under the EU AI Act, the open-source exemption from technical-documentation requirements does not apply to models classed as carrying systemic risk — which a frontier release like Spark would be. Publishing the weights would therefore not relieve Meta of those obligations in Europe, whichever license it chooses.

Meta Superintelligence Labs

Muse Spark's release marks the debut of Meta Superintelligence Labs — a new organization within Meta led by Alexandr Wang, the former CEO of Scale AI. Meta acquired Wang's involvement through a $14.3 billion deal with Scale AI in June 2025 — one of the largest AI talent acquisitions in history.

The creation of a dedicated "Superintelligence Labs" signals Meta's escalating ambition in AI, and gave the company a proprietary frontier model to power its consumer products directly — something the open-weight Llama line was never designed to do.

For most of 2026 that read as Meta walking away from open weights altogether. It did not turn out that way. On August 10, 2026 Meta released Muse Glimmer, a 30 billion parameter open-weight model, under a plain Apache 2.0 license — more permissive than any Llama release — alongside a 6,500-word essay in which Mark Zuckerberg argued that concentrated superintelligence is itself the safety risk and recommitted Meta to open releases.

For a few weeks that settled into a clean two-track story: a closed flagship for consumer scale, an open line for adoption. Zuckerberg's September 2 promise of open weights for Muse Spark itself unsettles it again. The boundary between the two tracks is no longer the product line but the specific checkpoint — Meta plans to open 1.2 while shipping 1.3 closed, which makes "is Muse Spark open?" a question with no single answer. Until a version actually lands with a license attached, treat the tracks as blurred rather than separate.

Key Capabilities

Multimodal Input

Muse Spark processes multiple input types:

• Text — standard conversational and analytical capabilities
• Voice — spoken queries processed natively (not speech-to-text conversion)
• Images — visual understanding and analysis from uploaded or camera-captured images

Output is currently text-only, with additional output modalities expected in future updates.

Multiple Reasoning Modes

The model supports several operational modes:

• Fast mode — quick responses for simple queries
• Reasoning modes — deeper analysis for complex questions, similar to extended thinking in competing models
• Shopping mode — commerce-focused mode for product search, comparison, and recommendations across Meta's marketplace and partner integrations

Platform Integration

Muse Spark is designed to power Meta AI across Meta's ecosystem:

• Meta AI app and website — standalone access (available at launch)
• WhatsApp — AI assistant in the world's most-used messaging platform (2 billion+ users)
• Instagram — creative AI features, content understanding
• Facebook — feed interaction, content discovery, Marketplace
• Messenger — conversational AI assistant
• Ray-Ban Meta smart glasses — voice-activated AI with visual understanding through the glasses' camera

This distribution across Meta's platforms gives Muse Spark potential access to over 3 billion monthly active users — the largest AI distribution channel in the world.

Muse Spark vs. Llama

The model lines serve different strategic purposes: the open releases commoditize the AI model layer, which benefits Meta by reducing competitors' moats, while Muse Spark gives Meta's own products a proprietary advantage. Muse Glimmer is the current expression of the open half — smaller than Llama 4, but licensed far more freely.

Llama 4 Behemoth Status

Llama 4 Behemoth — Meta's largest planned open-weight model (288 billion active parameters, approximately 2 trillion total) — has never been released. Multiple delays through 2025 pushed it back indefinitely and it was effectively deprioritized in favor of Muse Spark and the Meta Superintelligence Labs roadmap. Meta's return to open weights arrived through the much smaller Muse Glimmer instead, so Behemoth should be treated as shelved rather than pending.

Meta's AI Investment

Meta's AI capital expenditure for 2026 is projected at $115 to $135 billion — among the largest AI infrastructure investments by any company. This funds:

• Training infrastructure for Muse Spark and future models
• The expanded data center footprint required for inference at Meta's scale
• Research at Meta Superintelligence Labs, FAIR, and GenAI teams

Company Details

Related Tools

• Muse Glimmer — Meta's Apache 2.0 open-weight model for local agents (30 billion parameters)
• Llama 4 Maverick — Meta's open-weight frontier model (MoE, 400 billion total parameters)
• Meta AI — Meta's free AI assistant across its platforms
• ChatGPT — OpenAI's competing consumer AI product
• Claude Cowork — Anthropic's desktop agent for knowledge work

Key Takeaways

• Muse Spark is Meta's flagship model — the first from Meta Superintelligence Labs, led by Alexandr Wang (former Scale AI CEO, hired via a $14.3 billion deal)
• Muse Spark 1.1 (July 9, 2026) pushed Meta into the AI-coding race — enterprise-grade agentic coding, parallel subagents, and a one-million-token context, shipped in public preview through the new OpenAI-compatible Meta Model API
• Muse Spark 1.3 (September 2, 2026) is the current release and targets agentic reliability rather than raw capability — asking clarifying questions, calling for help when stuck, and confirming before consequential actions, which cuts the turns and tokens a task consumes
• Alexandr Wang calls 1.3 competitive with Claude Fable 5.1 and better than GPT-5.6 Sol at coding; Artificial Analysis independently puts it level with GPT-5.6 Sol, Claude Opus 5 and Grok 4.6 High on overall intelligence, a four-point gain over 1.2 — the coding claim specifically remains unverified
• Muse Spark ships cloud-served through the Meta Model API — designed to power Meta AI across WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban glasses, with potential reach of over 3 billion users
• The model accepts voice, text, and image inputs and features multiple modes including fast, reasoning, and shopping modes
• Whether Muse Spark gets open weights is unresolved. Zuckerberg promised them "soon" on September 2 without naming a version; Meta plans to publish 1.2 and has not decided on 1.3. Do not plan a deployment around them, and read the license when they land rather than assuming Muse Glimmer's Apache 2.0 carries over
• Meta runs two tracks rather than having abandoned open weights: Muse Glimmer (August 10, 2026) is a 30 billion parameter open-weight model under a plain Apache 2.0 license — more permissive than any Llama release — but the closed-versus-open boundary now runs between checkpoints of Spark itself, not cleanly between the two products
• Llama 4 Behemoth was deprioritized in favor of Muse Spark and should be treated as shelved; Meta's open-weight return came through the much smaller Glimmer instead
• Meta's 2026 AI capital expenditure of $115 to $135 billion is among the largest infrastructure investments by any company