Did Meta kill Llama? What Muse Spark means for its future

Did Meta kill Llama? What Muse Spark means for its future

The weights are still online. The frontier strategy that made them important may not be.

Meta did not announce the death of Llama. It announced something more careful.

When Muse Spark arrived in April, it came from Meta Superintelligence Labs, on a rebuilt training stack, under a new family name, without downloadable weights. Meta called it the first product of a ground-up overhaul of its AI effort. The company compared its efficiency directly with Llama 4 Maverick and said it could reach the same capability with more than an order of magnitude less compute.

That is not a routine successor. It is a replacement story with the word replacement removed.

Llama is not literally dead

The narrow answer is no. Existing Llama models are still downloadable, deployable, and useful. Meta told VentureBeat that its current Llama models would continue to be available as open source. A model does not stop working because its maker changes the launch deck.

But that statement only protects the files already released. It does not promise a Llama 5, a new frontier checkpoint, or continued investment at the level that made Llama the default American open-weight family. Meta answered the availability question and left the roadmap question open.

So there are three different meanings of “dead”:

• As software: no. Llama 3 and Llama 4 deployments keep running.
• As a supported ecosystem: not yet. Hosting, fine-tuning, and community work continue.
• As Meta's frontier strategy: the evidence now points to Muse, not Llama.

The third meaning is the one developers should budget around.

Why Meta needed a clean name

Llama 4 did not give Meta the reset it needed. Its April 2025 launch drew mixed reviews, and the benchmark presentation became part of the story after Meta used an unreleased experimental chat variant in its LMArena comparison. The technical debate mattered, but the brand damage mattered more: every new score came with an argument about which model had actually produced it.

Then Meta reorganized. Meta Superintelligence Labs brought a new leadership structure, a large recruiting push, new infrastructure, new architecture, and new data pipelines. Calling the result “Llama 5” would have tied that reset to the line Meta was trying to move past.

Muse does three jobs at once:

• It separates the new models from the Llama 4 launch.
• It moves Meta from an open-weight identity to an API and product identity.
• It gives “personal superintelligence” a family name that can sit inside Meta AI, glasses, Instagram, and developer tools.

That makes Muse Spark a rebrand, but not only a rebrand. The model family, deployment model, organization, and commercial route all changed with the name.

The closed turn is the real break

Llama made Meta unusual among the largest American labs. Developers could download the weights, run them on their own infrastructure, fine-tune them, and build without sending every prompt back to Meta. The license was not conventional open source, but the operational freedom was real.

Muse Spark reverses that relationship. Muse Spark 1.3 is available through Muse Code and Meta Model API. The standard API tier says customer data is not used to improve Meta's products; the dramatically cheaper Contributor tier is explicitly marked as data that is used to improve them. The model is something developers access, not something they possess.

That may be commercially sensible. Closed access lets Meta iterate faster, control safety boundaries, collect usage data, and charge for inference. It also removes the reason many developers adopted Llama in the first place.

An API-compatible Muse model can replace another API. It cannot replace a model running inside a hospital, government network, private repository, or local device where the weights are the product.

Is this a rebrand after failure?

Yes, with one qualification: rebranding does not make the underlying improvement fake.

The original Muse Spark was a large jump over Llama 4 Maverick in reported reasoning, multimodal work, and token efficiency. The family moved from a private preview in April to a public developer API, Muse Code, and version 1.3 by September. Meta is shipping quickly enough that Muse is already more than a press-release escape hatch.

The skeptical reading is still reasonable. Meta took the embarrassment attached to Llama 4, placed the next system under a new lab and a new name, and stopped leading with openness. It kept the useful part of the old story—scale, distribution, and developer interest—while changing the bargain.

That is why “Llama is still available” is not a complete defence. A museum is full of things that remain available.

What would prove Llama is alive

Meta can settle this without another interview. Ship one of these:

• A genuinely new Llama generation with competitive open weights.
• An open-weight Muse release with a clear license and a repeatable roadmap.
• A public commitment that separates the closed frontier line from an actively funded open line.

Until then, existing Llama models should be treated as durable artifacts, not evidence of future direction. Teams that need self-hosting should compare the models available now from the wider open ecosystem instead of planning around an unnamed future Meta release.

Muse Spark may become an excellent proprietary platform. It may even restore Meta to the frontier. Neither outcome keeps the Llama strategy alive.

The honest verdict is narrower than “Llama is dead” and harsher than “nothing changed”: Llama still exists, but Meta has stopped showing that its best new work will be Llama. Muse Spark is the new bet. Openness is now a promise Meta can revive later, not the product it is shipping today.

For code review, that distinction matters. ThinkReview lets teams compare models on the same real pull request across GitHub, GitLab, Azure DevOps, and Bitbucket. A new family name is easy. Better findings on your code are the evidence.

Sources: Meta's Muse Spark introduction, the Muse Spark 1.3 release, and VentureBeat's reporting on Meta's Llama response.