Meta's Muse Spark: The End of Open-Weight Llama

Meta introduced Muse Spark on April 8, 2026, the first model from its new Meta Superintelligence Labs, and the clearest break yet from the strategy that defined Meta’s AI for three years. Muse Spark is proprietary, with no open weights, no download link, and no Hugging Face card. For the company that made open-weight frontier models a movement with Llama, that’s the headline.

A New Lab, a New Direction

Muse Spark (code-named “Avocado”) is the debut of Meta Superintelligence Labs, the division Meta stood up to chase what it calls “personal superintelligence,” led by former DeepMind researcher and Llama team lead Joelle Pineau. The naming matters here, because this isn’t Llama 5 and it isn’t meant to be. Meta’s frontier development has moved entirely to the new Muse series, and the expectation across the industry is that Llama gets maintenance from here, not the investment that pushed it during 2023–2025.

What Muse Spark Does

It’s a natively multimodal reasoning model (visual and text in one context) with built-in tool use and multi-agent orchestration. The feature Meta leads with is Contemplating mode, which spins up multiple agents that reason in parallel before converging on an answer, aimed squarely at the extreme-reasoning tiers from competitors like Gemini Deep Think and GPT Pro.

The efficiency claim is the part that should make rivals pay attention: Meta says Muse Spark reaches the same capabilities with over an order of magnitude less compute than Llama 4 Maverick, its previous mid-size flagship. If that holds up, the benchmark line matters less than the cost curve underneath it.

The Benchmarks

Meta reports that Contemplating mode hits 58% on Humanity’s Last Exam and 38% on FrontierScience Research, and it posts a category-leading 42.8 on HealthBench Hard (open-ended health queries), where it’s substantially ahead of competitors. On the broader Artificial Intelligence Index v4.0 it scores 52, which places it fourth overall behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6.

As always, these are vendor-selected figures run on benchmarks the vendor chose to highlight. Fourth on the composite index, first on a domain like health: that’s a model that’s strong without being the outright leader, which tracks with Meta entering 2026 a step behind the frontier and spending heavily to close the gap.

Closed, Not Open

This is the decision developers will argue about. Llama’s open weights built an enormous ecosystem: fine-tunes, quantizations, local deployments, an entire stack of tooling that assumed you could run the model yourself. Muse Spark walks away from all of it. Meta has said it “hopes to open-source future versions,” but for now the company’s most capable model is a closed API, the same posture as OpenAI and Anthropic.

For teams that chose Llama specifically because it was open (data residency, on-prem requirements, no per-token cost), Muse Spark doesn’t replace it. Those teams are now looking harder at the open-weight frontier elsewhere, which in 2026 increasingly means DeepSeek V4 or the open Qwen line, both of which we rate in our best open source models for coding roundup.

What It Means

Muse Spark is available at meta.ai and in the Meta AI app, with a private API preview for select developers. The capability is real and the efficiency claims are genuinely interesting, but the strategic shift is the story: Meta, the standard-bearer for open-weight AI, has put its frontier behind a closed door. Whether future Muse models reopen it is the question worth watching.