Meta Platforms Inc. has turned its back on the open-source darling that powered developers’ dreams. Llama, once hailed with 1 billion downloads by March 2025 and climbing to 1.2 billion soon after, now sits on the sidelines. Meta’s own blog trumpeted those figures. But fast-forward to April 2026. The company unleashes Muse Spark—a proprietary powerhouse from its new Meta Superintelligence Labs. No weights to download. No self-hosting. Cloud-only access via private API preview.
This shift stings. Developers who built empires on Llama’s open weights face a void. Mark Zuckerberg, once proclaiming “Open Source AI is the Path Forward” in October 2024, now chases profits through closed doors. Why? Llama lagged. Llama 4, released in April 2025, flopped against rivals like ChatGPT and Claude. Benchmarks exposed the gap. Internal frustration boiled over.
Enter Alexandr Wang. Meta shelled out $14.3 billion for a 49% stake in Scale AI last June, installing Wang as chief AI officer. He leads Meta Superintelligence Labs. In nine months, his team scrapped Llama’s foundation. New infrastructure. Fresh architecture. Custom data pipelines. The result: Muse Spark, announced April 8, 2026, on Meta AI’s blog. “We rebuilt our AI stack from scratch,” Wang posted on X.
Muse Spark packs multimodal punch—text, images, tools. It powers Meta AI across apps and glasses. Benchmarks? Solid. Scores 52 on the Artificial Analysis Intelligence Index, trailing only Gemini 3.1 Pro and GPT-5.4 at 57, and Claude Opus 4.6 at 53, per Artificial Analysis. Dominates HealthBench Hard at 42.8%, trouncing Gemini’s 20.6% and Claude’s 14.8%, as detailed in Fortune. GPQA Diamond: 89.5%. LiveCodeBench Pro: 80%. Yet coding lags—59% on Terminal-Bench 2.0 versus GPT-5.4’s 75.1%.
Llama Loyalists Left Stranded
No migration path exists. Llama’s downloadable weights clash with Muse Spark’s API-only model. “There is no migration path from Llama to Muse Spark because they have fundamentally different deployment models,” writes Steven J. Vaughan-Nichols in The New Stack. Current Llama models stay available. Expect patches, not breakthroughs. Meta confirms: they’ll linger on cloud providers. But frontier cash flows to Muse.
Andrew Ng nails it in his newsletter The Batch: “The proprietary release has raised concerns among developers, many of whom have built projects on open-weights Llama models.” Ng sees upside for Meta’s business chase against OpenAI, Google, Anthropic. Downside? Developers lose their U.S. open-weight champion.
Options? Stick with aging Llamas. Switch to Mistral, DeepSeek, Qwen. Or pay for proprietary APIs—Meta’s own teams ditched Llama for Claude Sonnet pre-Spark. Costs mount. Rework APIs. Retrain data. Rebuild tools. Thousands of firms scramble.
Forks offer lifelines. Llama.cpp, the C++ inference king, spans models beyond Meta’s flock. Ik_llama.cpp boosts CPU/GPU speed. Rkllama taps Rockchip NPUs for embeds. Llama-rs races in Rust. OpenLLaMA reproduces originals under Apache license—3B to 13B params, trillion-token trained.
Meta’s High-Stakes Wager Pays Early Dividends
Zuckerberg isn’t flinching. Q1 2026 earnings bumped AI capex to $115-135 billion, nearly double last year’s, per The Wall Street Journal. Muse Spark rolls to WhatsApp, Instagram, Facebook, Messenger. Free at meta.ai. Contemplating mode—multi-agent reasoning—deploys gradually. Safety report boasts bias resistance, risk mitigations, as in Meta’s scaling blog.
But cracks show. Agentic tasks trail: GDPval-AA ELO at 1427-1444, behind GPT-5.4’s 1672 and Claude’s 1607. Abstract reasoning? ARC-AGI-2 at 42.5% versus Gemini’s 76.5%. X buzz mixes hype and shade. “Zuck is officially the most evil man in AI,” quips one user. Others praise Spark’s edge over Gemini in agents with calendar access.
Meta bets big on personal superintelligence. Muse Spark eyes your camera, glasses—real-world sight. Llama built the ecosystem. Spark locks the gate. Developers adapt or perish. Meta surges ahead. Profit motives trump open ideals. The AI arms race demands no less.
And so the board resets. Forks thrive. Rivals pounce. Watch Q2 earnings. Capex burns. Benchmarks evolve. Muse Spark’s true test? Billions of users, not leaderboards.
WebProNews is an iEntry Publication