Muse Spark: Meta Spent $14 Billion to Catch Up in AI, and Ditched Open Source Along the Way

After a year of near-silence and a Llama 4 launch widely considered a flop, Meta just dropped Muse Spark, its first proprietary AI model. Nine months of work, $14.3 billion poured into Scale AI, and a talent-raiding spree across OpenAI, Anthropic, and Google. The model is free on meta.ai, but you'll need your Facebook or Instagram login to get in.

What Muse Spark actually brings to the AI table

Muse Spark is a natively multimodal reasoning model that accepts text, voice, and image inputs. On the Artificial Analysis Intelligence Index, it scores 52, placing it behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6, but miles ahead of last year's Llama 4 Maverick (18) and Scout (13). That jump is hard to ignore. Meta does admit, however, that the model lags behind on coding workflows and long-horizon agentic tasks.

Three modes are available. Instant handles quick questions, Thinking tackles multi-step reasoning, and a forthcoming Contemplating mode will orchestrate multiple agents in parallel. Meta says Muse Spark achieves the same performance as Llama 4 Maverick with an order of magnitude less compute. Whether that efficiency claim holds up under independent testing remains to be seen.

Meta ditches open source for Muse Spark

Here's where it gets interesting. Muse Spark is Meta's first model that is not released as open weights. No download, no tinkering, no fine-tuning on your own hardware. For a company whose Llama ecosystem had reached 1.2 billion downloads by early 2026, this is a genuine U-turn. Zuckerberg promises open-source Muse models later, but with no firm timeline.

The developer community that made Llama successful is now left waiting. Meanwhile, Muse Spark will roll out across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta AI glasses in the coming weeks. Meta's bet is clear: reach its 3 billion users first, worry about third-party developers second.

Privacy concerns around Muse Spark

You can't talk about a Meta product without talking about data. Muse Spark requires a Meta account login, and the company's privacy policy sets few limits on how data shared with its AI can be used. Meta hasn't explicitly confirmed it will tap your Facebook profile data, but the company has a history of training on public user content. A "personal superintelligence" product attached to your social graph should make anyone pause.

The Shopping mode makes the point even clearer. Style tips, gift suggestions, room decor ideas, all drawn from creator content and the people you follow. Useful or creepy? Depends on how you feel about Meta knowing what you want before you do.

How Muse Spark compares to ChatGPT, Gemini, and Claude

Let's be real: Muse Spark doesn't top any leaderboard outright. Meta admits this openly, which is honestly refreshing after last year's Llama 4 benchmark scandal. The model is competitive in multimodal vision and health tasks, where it outperforms GPT-5.4 and Gemini 3.1 Pro on HealthBench Hard. On coding and agentic benchmarks, Claude, GPT-5.4, and Gemini all pull ahead.

Meta stock jumped over 9% on launch day, so Wall Street seems convinced the billions produced something real. The bigger question isn't whether Muse Spark is good. The question is whether Meta can monetize it. The paid API is coming, but the competition has a year's head start. Read the full technical details on Meta's AI blog.