Meta AI
Pro
~7 min read
New · July 2026
Muse Spark 1.1: the brain behind Meta AI just got ambitious.
On July 9 — two days after Muse Image — Meta Superintelligence Labs shipped Muse Spark 1.1, the model that powers Thinking mode in the Meta AI app. It's built for agentic work: planning multi-step tasks, using tools, operating a computer, and keeping track of huge amounts of context. And for the first time ever, Meta opened a developer API for one of its frontier models. Here's what actually changed, minus the press-release fog.
Update — August 5, 2026: Meta shipped
Muse Spark 1.2, a coding-focused revision of this model, as the engine behind
Muse Code — Meta's first terminal coding agent, launched the same day in beta with a $0.30/1M-token "contributor tier" that trains on your code. The launch, the tool, and the pricing trade are covered in
Muse Code: Meta's coding agent — and the 30-cent trade.
Update — August 10, 2026: Meta released
Muse Glimmer — and put the weights up under a permissive
Apache 2.0 licence, which partly reverses the closed posture described further down this page. Glimmer is a
30-billion-parameter agentic model
distilled from Muse Spark: Meta trained it on Spark's own outputs, then tuned it for long-horizon tool use. The point is where it runs. Compressed to roughly 4-bit, the model drops
under 20 GB and fits — with room for its image encoder and a speed-up drafter — inside a
24 GB or 32 GB budget: a MacBook M4/M5 Max, or one consumer GPU like an RTX 5090. Weights are on Hugging Face now, with Ollama, LM Studio, llama.cpp, MLX and ExecuTorch support promised "in the coming days," and it is built to plug into agent scaffolds including OpenClaw. So Meta now runs both plays at once: a paid API for the frontier model, open weights for the small one you can keep offline. What "open weights" actually buys you is in
Inkling: what it means when a lab gives away the model.
What shipped, in one paragraph
Muse Spark 1.1 is a multimodal reasoning model built for agentic tasks, with what Meta calls major gains in tool use, computer use, coding, and multimodal understanding. It's live now in Thinking mode in the Meta AI app and on meta.ai — nothing to install, no waitlist. Alongside it, Meta launched the Meta Model API in public preview, so developers can build on the model directly for the first time.
The announcement, decoded
Model launches come wrapped in jargon. Here's Meta's language translated into what it means when you're the one typing:
Meta says
What it means for you
"Built for agentic tasks"
You can hand it a goal ("plan this," "organize that," "find and compare these") instead of a single question, and it works through the steps itself.
"1 million token context, actively managed"
It can hold roughly a dozen novels' worth of conversation and documents in mind — and it decides what to keep and what to compress, so long sessions don't fall apart.
"Computer use"
The model can operate apps and a browser on your behalf — clicking, typing, and switching between programs — and it chooses when to automate with a script versus click like a human.
"Orchestrates multi-agent systems"
On a big job, it splits itself into a coordinator plus parallel helpers, the way a project lead delegates to a team. You just see faster results.
"Multimodal understanding"
It reads images, video, and audio as fluently as text — and can act on what it sees, not just describe it.
Meta's own demo, worth knowing about
The Marketplace agent. In Meta's launch demo, a user shoots a quick smartphone video of an item, and Muse Spark 1.1 pulls the useful frames as photos, reasons about what the product is, then operates the browser itself to create the Facebook Marketplace listing. That's the shape of what "agentic" means in practice: perception, judgment, and action in one loop — not a chatbot with better answers.
The bigger story: Meta is selling model access now
For years Meta's play was the opposite of OpenAI's and Anthropic's: give the models away as open-weights Llama and win on ubiquity. Muse Spark 1.1 breaks that pattern. No open weights shipped with it — you reach the model through Meta's apps or through the new Meta Model API, a commercial product now in public preview. (The pattern is now a split, not a reversal: on August 10 Meta open-sourced Muse Glimmer, a small model distilled from Spark, under Apache 2.0 — the frontier model stays paid, the laptop-sized one is free.) Launch coverage widely notes the API undercuts rival frontier models on price; check Meta's developer site for current rates rather than trusting any article's snapshot, ours included.
Why you should care even if you never touch an API: paid access means Meta now has a direct revenue reason to keep this model sharp, and early partners — including coding tools like Cline and Replit and the OpenClaw agent platform — are already wiring it in. Expect Muse Spark to start showing up in tools that have nothing to do with Meta's apps.
Try it today — a 2-minute test drive
Open the Meta AI app (or meta.ai), switch the conversation to Thinking mode, and give it a genuinely multi-step task instead of a question. A good first test: "Plan a garage sale for the last Saturday of this month — make me a checklist, a pricing guide for typical items, and a short listing I can post." Watch how it structures the work rather than dumping one long answer. That structured, plan-first behavior is the 1.1 difference.
The honest part
"Public preview" means moving parts. The API is days old, and Meta AI features have a track record of changing fast — Muse Image lost its @-mention feature three days after launch. The capability claims above are Meta's own evaluation numbers; independent testing takes weeks to catch up. And computer use inside consumer Meta AI is arriving unevenly — what you can trigger from the app today may differ from the demos. Treat this release as a direction, not a finished promise.
Muse Spark 1.1 — quick answers
What is Muse Spark 1.1?
Meta Superintelligence Labs' July 9, 2026 model — a multimodal reasoning model built for agentic tasks, with major gains in tool use, computer use, coding, and multimodal understanding. It powers Thinking mode in Meta AI.
Where can I use it?
In Thinking mode in the Meta AI app and on meta.ai, available now. Developers can build with it via the new Meta Model API, in public preview.
Is it free?
Inside Meta AI, yes — it's part of the assistant. The developer API is a commercial product; current pricing is on Meta's developer site.
Is it open source like Llama?
Not Muse Spark 1.1 itself — no open weights were published, so access is through Meta's apps or the paid API. But Meta has not abandoned open weights: on August 10, 2026 it released Muse Glimmer, a 30B model distilled from Spark, under Apache 2.0 on Hugging Face. Frontier model paid, small model free.
What is Muse Glimmer?
Meta's 30-billion-parameter open-weights agentic model, released August 10, 2026 under Apache 2.0 and distilled from Muse Spark. Compressed to about 4-bit it needs under 20 GB, so it runs locally on a Mac or PC with a single consumer GPU — built for always-on local agents, tool calling, local coding and image-plus-text input.