According to Meta, on April 8, 2026, Meta Superintelligence Labs (MSL) officially launched Muse Spark — Meta's first in-house large language model (LLM), used to power Meta AI across the company's products. The striking part: instead of continuing the Llama line's open-source tradition, Muse Spark takes a proprietary (closed) path — at least in its early stage.
Quick summary
- When: launched April 8, 2026; with a major update on May 12, 2026 (per Meta).
- What: Muse Spark — Meta's first in-house LLM, built by Meta Superintelligence Labs.
- Direction: proprietary — initially available only via API as a private preview for select partners.
- Positioning: multimodal reasoning, strong at visual coding and perception, able to handle complex questions in science, math and medicine (per Meta).
- The news angle: pressure is mounting on Zuckerberg to commercialize after the enormous investment in his superintelligence ambitions.
What happened?
According to Meta's official announcement, Meta Superintelligence Labs launched Muse Spark on April 8, 2026, with a major follow-up update on May 12, 2026. This is the first large language model developed in-house by Meta within this new lab, and it is used to power the Meta AI assistant across the company's products.
The most notable move lies in the distribution model: rather than releasing open weights as the Llama line did previously, Meta is steering Muse Spark toward a proprietary (closed) format — initially granting access only via API as a private preview for a handful of select partners. According to Meta, the company "hopes to open-source later versions," but for now this is a temporary departure from the open-source tradition that has been Llama's signature.
How is Muse Spark positioned?
According to Meta, Muse Spark is positioned as a multimodal reasoning model, with strengths in visual coding (programming based on images/interfaces) and perception (visual understanding). The company says the model is capable of handling complex questions in fields such as science, math and medicine.
In other words, this is not merely a Llama upgrade but a product that reflects Meta Superintelligence Labs' new direction: focusing on deep, multimodal reasoning capabilities to compete head-on with other frontier models on the market.
Why did Meta choose the proprietary path?
This is the most surprising point, because for years Meta has been the face of the open-weight model movement with its Llama line. The fact that Muse Spark starts out closed — opening its API only to select partners — shows that even one of the biggest champions of open source is reconsidering its strategy as it enters the "superintelligence" race.
Meta says it still "hopes to open-source later versions." For now, however, this is a commitment about the future rather than a present reality — and any business building on the expectation that "Meta is always open" should take note of this shift.
Commercialization pressure on Zuckerberg
According to CNBC (June 14, 2026), after the enormous investment in superintelligence ambitions — including bringing Alexandr Wang to Meta (through a deal involving Scale AI) — pressure is mounting on Mark Zuckerberg to commercialize Muse Spark and prove the value of that investment.
This partly explains why Meta chose a controlled distribution approach via API for partners: a proprietary model is easier to commercialize and to protect a competitive edge than releasing open weights for free.
Lessons for Vietnamese businesses
Meta's pivot is an important signal for every business building products or internal assistants on top of external AI:
- "Open" is not a permanent commitment. A provider once famous for open source can still go closed when its interests change.
- Private-preview API = conditional access. If you are not on the list of select partners, the latest capabilities may be out of reach.
- Depending on a single closed provider is a cost and availability risk: pricing, terms and access are all decided by them.
When even Meta is shifting toward proprietary, businesses should give all the more thought to open-source models running in-house (on their own infrastructure) to stay in control of access, cost and data — rather than relying entirely on the decisions of a foreign provider.
FAQ
Is Muse Spark open-source?
According to Meta, in its early stage Muse Spark takes a proprietary (closed) path, available only via API as a private preview for select partners. Meta says it "hopes to open-source later versions," but this is a future expectation, not a present reality.
How is Muse Spark different from Llama?
Llama is Meta's line of open-weight models. Muse Spark is Meta Superintelligence Labs' first in-house LLM, positioned as a multimodal reasoning model and going proprietary — a clear difference in distribution strategy (per Meta).
Can Vietnamese businesses use Muse Spark?
At the time of writing, Muse Spark is only available via API as a private preview for select partners. Businesses outside this list most likely cannot access it yet. This is reference information and may change.
Take charge of AI, without provider lock-in
When even Meta is going closed, Namtech helps businesses deploy open-source models running in-house — staying in control of access, cost and data, without depending on a foreign provider's decisions.
Book a free consultationNote: This article is compiled from public sources as of 22/06/2026; the situation may change. For reference only, not investment or legal advice.