DEV Community

jidonglab
jidonglab

Posted on • Originally published at spoonai.me

Meta Ditched Llama for a Closed Model Called Muse Spark — Open Source AI Just Lost Its Biggest Champion

In 2024, Mark Zuckerberg was open source's biggest cheerleader.

He released Llama 2 to the world, declared that hoarding AI was wrong, and set new standards with Llama 3. Millions of developers built on Meta's open-weight models. Meta was hailed as the champion of AI democratization.

Then, in April 2026, Meta went the other direction entirely. The company launched Muse Spark, its first fully proprietary AI model. Available only through the Meta AI app and website, with API access limited to hand-picked partners. No open weights. No community downloads. More closed than OpenAI or Anthropic.

To Understand Muse Spark, Start With Llama 4's Stumble

The backstory matters. Llama 4 launched in early 2026 to impressive specs but disappointing market reception.

Scout had 109B parameters with a 10-million-token context window. Maverick packed 400B parameters. Technically excellent. But the market yawned. The problem wasn't the models themselves. It was the economics.

Model Parameters Context Benchmark Market Response
Llama 4 Scout 109B (17B active) 10M tokens Strong Lukewarm
Llama 4 Maverick 400B 1M tokens Top-tier Moderate
GPT-5 Turbo Undisclosed Undisclosed Top-tier Hot
Claude Opus 4 Undisclosed 200K SWE-bench 72.1% Hot

Meta had assumed open-sourcing models would drive developers toward Meta's platforms. Instead, AWS, Azure, and Google Cloud hosted Llama and captured the revenue. Meta was spending billions on training while cloud providers pocketed the margins.

That realization hit Zuckerberg hard. In summer 2025, he decided to overhaul Meta's entire AI organization.

Enter Alexandr Wang and MSL

Zuckerberg's pick to lead the transformation was Alexandr Wang, the 29-year-old co-founder and former CEO of Scale AI.

Wang co-founded Scale AI at 19 and grew it to a $14B valuation. He understood data quality better than almost anyone in the industry. Meta brought him in through a jaw-dropping $14B acquisition of Scale AI, then gave him the keys to a brand-new division: Meta Superintelligence Labs (MSL).

MSL operates independently from FAIR (Fundamental AI Research), Meta's longtime open-science arm. Where FAIR published papers and released models freely, MSL is laser-focused on commercial competitiveness. Muse Spark is MSL's debut product.

What Makes Muse Spark Different

Here's the deal: Muse Spark is locked inside Meta's walls.

You can use it on the Meta AI app and website. Select partners get an API preview. That's it. No weights to download. No self-hosting. No fine-tuning on your own data. This is more restrictive than ChatGPT (which at least has a broadly available API) or Claude (which offers enterprise API access to anyone).

Meta's logic is transparent: make Muse Spark the killer feature that keeps 3 billion Facebook, Instagram, and WhatsApp users inside Meta's ecosystem. Not a platform play for developers. A retention play for consumers.

Meta went from "champion of open source AI" to "the most closed AI company" in just 18 months.

The Bigger Picture — Open Source AI Enters a Multi-Polar Era

Does Meta's exit kill open-source AI? Not necessarily.

The landscape was already diversifying. By late 2025, Chinese models from Alibaba and DeepSeek accounted for 41% of downloads on Hugging Face. In April 2026, Google shipped Gemma 4 under Apache 2.0, Zhipu AI released GLM-5.1 under MIT, and numerous smaller labs are producing competitive open models.

Open Model Origin Parameters License
Gemma 4 Google 27B dense, 26B MoE Apache 2.0
GLM-5.1 Zhipu AI (China) 744B MoE (40B active) MIT
Qwen 3 Alibaba (China) Various sizes Apache 2.0
DeepSeek V3+ DeepSeek (China) 671B MoE Custom

There's an irony worth noting: while the U.S. government blocks China's access to AI chips, Chinese companies are expanding their influence over the global developer ecosystem through open-source software. Meta's retreat creates a vacuum that Chinese labs are eagerly filling.

What This Means for Developers

Three things to pay attention to.

First, if you've been building on Llama, start evaluating alternatives now. Meta says existing Llama models will remain available, but there's no guarantee the next generation will be open. Gemma 4, GLM-5.1, and Qwen 3 are realistic alternatives.

Second, the "open source equals free lunch" illusion is cracking. Training frontier models costs billions. Giving that away indefinitely was never a sustainable business model. Expect a future where small models stay free while frontier models go behind paywalls.

Third, the real power in AI is shifting from model developers to platforms. Just as Meta is locking Muse Spark inside its apps, the long-term winner won't be whoever builds the best model. It'll be whoever controls the distribution.


References


Originally published on spoonai.me | Daily AI briefing at spoonai.me

Top comments (0)