Meta Muse AI: 3 models in 3 weeks — Spark 1.3 (Sept 2), Glimmer (Aug), Muse agent (Sept 8)

Meta launched three Muse-branded AI products in August-September 2026: Muse Spark 1.3 (Sept 2, closed API model for developers), Muse Glimmer (August, open-weight variant for consumer hardware), and Muse agent (Sept 8, the world's first personal AI agent built for everyone, integrated across Facebook/Instagram/WhatsApp/Meta AI). Muse Spark 1.3 beats OpenAI's GPT-5.6 Sol on coding benchmarks (75.4% DeepSWE 1.1, 88.8% Terminal-Bench 2.1) and matches Anthropic's Claude Opus 5. A fourth model, codenamed "Watermelon," is in development. This is Meta's strategic pivot from open-source Llama to a closed, paid, API-gated lineup — structurally similar to GPT-4o and Claude. Chief AI Officer Alexandr Wang announced all three products.

Quick answers: What are Meta's three Muse AI products?

Product	Launch	Audience	Distribution
Muse Spark 1.3	Sept 2, 2026	Developers	Closed API + Muse Code (paid)
Muse Glimmer	August 2026	Device makers / open-source community	Open weights, downloadable
Muse (agent)	Sept 8, 2026	Consumers across Facebook, Instagram, WhatsApp, Meta AI	Built into Meta apps

A fourth model, Watermelon, is in development — no release date confirmed.

What is Meta Muse AI?

Meta Muse AI is the umbrella name for Meta's 2026 AI model family, succeeding the Llama 4 series. The Muse lineup spans four tracks: Spark (the most capable paid API model), Glimmer (open-weight variant for devices), Muse (consumer personal AI agent), and Watermelon (next-gen flagship, in progress). The umbrella rebrand signals Meta's pivot from open-source-first (Llama) to closed-API-first (Muse Spark), with the open-source model (Muse Glimmer) surviving as a separate track for device and research use.

The rebrand also marks Meta's first model lineup under Chief AI Officer Alexandr Wang, who joined Meta in mid-2025 from Scale AI. Wang has publicly positioned Muse as the model family that delivers "personal superintelligence" — hyper-personalized AI that knows each user's preferences, history, and context across Meta's apps.

What is Muse Spark 1.3?

Muse Spark 1.3 is Meta's most capable closed AI model, released September 2, 2026. It is the fourth iteration of the Spark line (1.0 in April 2026, 1.1 in July, 1.2 on August 5, 1.3 on September 2). Spark 1.3 is positioned for coding, agentic tasks, and long-context work.

Benchmark	Spark 1.3 score	Beats / matches
DeepSWE 1.1 (software engineering)	75.4%	Beats GPT-5.6 Sol
Terminal-Bench 2.1 (terminal / agentic)	88.8%	Matches Claude Opus 5
Long-context retrieval (256K-512K tokens)	98.5%	New SOTA
Long-context retrieval (512K-1M tokens)	98.1%	New SOTA
Token efficiency	-25% tokens vs Spark 1.2	Same task completion

Spark 1.3 is closed-source (no model weights released) and API-gated (paid developer access only). Meta opened paid API access on September 3, 2026. Pricing has not been disclosed publicly, but per-token rates are reportedly competitive with OpenAI's GPT-5.6 and Anthropic's Claude Opus 5.

How does Muse Spark 1.3 differ from Llama 4?

Muse Spark is closed-API and paid; Llama 4 was open-source and free. This is Meta's biggest strategic AI shift since releasing Llama 1 in 2023.

Dimension	Llama 4 (April 2025)	Muse Spark 1.3 (Sept 2026)
Weights	Open (downloadable)	Closed
Distribution	Hugging Face + self-host	Meta Model API + Muse Code only
Pricing	Free	Paid (per-token, similar to GPT-5.6)
Best use case	Research, self-hosting, fine-tuning	Production apps, agentic coding
SOTA benchmark	Llama 4 Maverick: 84.0% MMLU	Spark 1.3: 75.4% DeepSWE, 88.8% Terminal-Bench

The shift is strategic: Meta is no longer competing with OpenAI by giving away the model. Instead, Spark competes with GPT-5.6 and Claude Opus 5 on paid developer mindshare, while Muse Glimmer (open-weight) keeps the open-source community engaged.

What is Muse Glimmer?

Muse Glimmer is the open-weight variant of the Muse family, released August 2026 for consumer devices and research use. It is smaller and less capable than Spark 1.3 but can run locally on-device (smartphones, laptops, edge devices) without cloud connectivity.

Glimmer is positioned for:

• On-device AI: phones, laptops, AR/VR headsets
• Privacy-sensitive applications: no data leaves the device
• Research / fine-tuning: open weights allow community customization
• Lower-latency use cases: edge inference vs cloud round-trip

Meta has not released a specific benchmark scorecard for Glimmer vs Spark 1.3. Per the announcement, Glimmer trades 20-30% capability for ~10x lower compute footprint.

What is the Muse personal agent?

Muse (the agent, not the model family) is the world's first personal AI agent built for everyone, launched September 8, 2026 across Facebook, Instagram, WhatsApp, and Meta AI. Unlike a chatbot that waits for commands, Muse is proactive — it surfaces reminders, suggestions, and actions based on your behavior across Meta's apps.

Capabilities (per Meta's announcement):

• Cross-app memory: remembers context from Instagram DMs, WhatsApp chats, Facebook events
• Proactive reminders: surfaces "Your package ships tomorrow" or "Sarah's birthday is Friday" without being asked
• Multi-step agentic tasks: "Plan a trip to Tokyo next month" → books flights, hotels, restaurant reservations
• Voice + text: works in WhatsApp voice notes, Instagram Reels search, Facebook Marketplace chats

Privacy caveat: Muse reads your messages and behavior across all four apps to build its model of you. Meta has stated that user data is not used to train Muse's underlying model, but is used for the per-user agent memory. Users can disable Muse memory or delete it from settings.

What is Watermelon?

Watermelon is Meta's next-generation flagship model, currently in development. No release date has been confirmed. Per Meta's internal communications leaked in August 2026, Watermelon is positioned as the GPT-6-class competitor — a single multimodal model (text, image, audio, video) with reasoning capability beyond Spark 1.3.

If Watermelon ships on Meta's stated internal timeline, it could land in Q2-Q3 2027. Until then, Spark 1.3 is the production flagship.

How does Muse AI compare to GPT-5.6 and Claude Opus 5?

Dimension	Meta Muse Spark 1.3	OpenAI GPT-5.6 Sol	Anthropic Claude Opus 5
DeepSWE 1.1	75.4%	~70%	~76%
Terminal-Bench 2.1	88.8%	~85%	~89%
Long-context (256K-512K)	98.5%	~96%	~97%
Long-context (1M+)	98.1%	~94%	~95%
API cost per million	TBD	$15/$60 (input/output)	$15/$75
Open weights?	No	No	No
Multimodal?	Yes (text + image)	Yes (text + image + audio)	Yes (text + image)

Muse Spark 1.3 is the strongest pure-coding model in the comparison but trails GPT-5.6 in multimodal breadth (audio in particular).

Should I use Muse Spark, GPT-5.6, or Claude Opus 5?

Use case	Recommended
Coding agents / SWE-bench style work	Muse Spark 1.3 (best DeepSWE)
Terminal / agentic / DevOps	Tied: Muse Spark 1.3 vs Claude Opus 5
Long-context document analysis (1M+ tokens)	Muse Spark 1.3 (best retrieval)
Multimodal (audio + video)	GPT-5.6 Sol
Privacy / on-device	Muse Glimmer (open-weight, local)
Consumer agent (Facebook/IG/WhatsApp)	Muse (the agent)
Chatbot / conversational	GPT-5.6 (most general)

For most production use cases today, the right answer depends on the specific workload. Muse Spark 1.3 is the strongest coding / long-context option as of September 2026.

When did Meta's three Muse launches happen?

Date	Launch	Significance
August 2026	Muse Glimmer (open-weight)	Keeps open-source community engaged
Sept 2, 2026	Muse Spark 1.3 (paid API)	Marks Meta's closed-API pivot
Sept 8, 2026	Muse (personal agent)	Consumer rollout across Meta's apps
Q2-Q3 2027 (projected)	Watermelon (next-gen flagship)	Multimodal reasoning

---

Bottom line: Meta's 2026 Muse lineup marks the company's pivot from open-source-first (Llama) to closed-API-first (Spark), while keeping the open-source community alive via Glimmer. The Sept 8 personal-agent launch is the most consequential for consumers — Muse is now embedded in every Meta app. For developers, Spark 1.3 is worth evaluating if your workload is coding, agentic, or long-context. For on-device / privacy-sensitive use, Glimmer is the open-weight choice. For consumers, the Muse agent will start showing up in your Facebook/Instagram/WhatsApp/Meta AI apps within the next few weeks as Meta rolls it out globally.

For the full Meta AI coverage, see our Meta news hub. For the personal-agent rollout, see our Muse agent launch coverage. For a coding-benchmark deep-dive of Spark 1.3 vs GPT-5.6 vs Claude Opus 5, see our AI model benchmark comparison. For Apple silicon coverage (the chip side of this competition), see /tech/.