Key Highlights –
- Meta launched Muse Spark, its first model from Meta Superintelligence Labs, led by chief AI officer Alexandr Wang
- Muse Spark is a closed, proprietary model which is the exact opposite of Meta’s Llama open-source strategy
- It is already live in the Meta AI app and on meta.ai, and will arrive on WhatsApp, Instagram, Facebook, and Ray-Ban glasses within weeks
After spending 12 months rebuilding its AI operation from scratch, Meta unveiled its the first model from Meta Superintelligence Labs – Muse Spark. Labs was Mark Zuckerberg first division created in mid-2025 after Llama 4 faced poor reviews and an admission that benchmark results had been inflated using unreleased, fine-tuned model variants. Muse Spark is Meta’s attempt to make up for the lost goal.
The model was developed under the internal codename Avocado and is now live in the Meta AI app and on meta.ai.
What Muse Spark Is
Meta describes it as small and fast by design. However, what’s worth examining is Meta’s claim of its efficiency. According to Meta’s technical blog, Muse Spark achieves the same capability level as Llama 4 Maverick using over an order of magnitude less compute. The technique behind this is called thought compression. During reinforcement learning, the model is penalised for spending too many reasoning tokens on a problem. Hence, it learns to reach correct answers with fewer steps.
The result places Muse Spark at 52 on the Artificial Analysis Intelligence Index. That puts it in the top five globally benchmarked models. However, independent verification of that figure has not yet been published.
Two modes are available. Thinking handles complex queries such as legal documents, nutritional analysis from photos, health questions with charts. Contemplating is the deeper reasoning mode, positioned to compete with Gemini 3.1 Deep Think and GPT-5.4 Pro. Meta’s own benchmarks show Muse Spark Thinking competitive against Claude Opus 4.6 Max, Gemini 3.1 Pro High, GPT-5.4 Xhigh, and Grok 4.2 Reasoning.
The multimodal element is real and practical. Users can photograph a food label and ask about protein content or scan a product and compare it to alternatives. When Muse Spark arrives to the Ray-Ban Meta glasses in coming weeks, that same capability will be enabled into your field of vision.
How is Muse Spark Different from Meta’s Llama
Meta built its AI reputation on open-source. The Llama series gave researchers, developers, and startups access to capable models they could run, modify, and deploy freely. That openness created a global ecosystem. Billions of downloads. Thousands of applications built on top.
Muse Spark is none of that. It is a closed model, and its code is not public. Meta says it hopes to open-source future versions. For now, access is limited to the Meta AI products and a private API preview for select partners.
The stated reason is “personal superintelligence,” building an AI that understands your world because it is built on your data, your relationships, your context across Meta’s platforms. A closed model gives Meta control over how that context is used. An open-source model does not.
Mark Zuckerberg’s Goals with Muse Spark
Nine months ago, Meta paid $14.3 billion for a 49% non-voting stake in Scale AI and brought in its 29-year-old co-founder Alexandr Wang as chief AI officer. Wang and Zuckerberg then recruited researchers from OpenAI, Anthropic, and Google at reported packages running into hundreds of millions of dollars in equity.
Meta’s AI capital expenditure for 2026 is between $115 billion and $135 billion. That is nearly double last year’s figure.
Muse Spark is the first public output from that investment. Meta stock rose nearly 9% on the announcement, although the broader market moved sharply upward on the same day following news of a pause in Iran strikes. Interestingly, the next Muse generation is already in development.
You can also read – WhatsApp Will Finally Introduces Noise Cancellation in Video & Voice Calls
Wrapping Up
Whether Muse Spark’s benchmarks hold up to independent testing is the open question. Meta has been caught inflating benchmark results before. That history means independent verification matters more here than it would for a first-time offender.
If the numbers are real, Meta is back in the race. If they are not, this becomes the second consecutive AI announcement Meta cannot stand behind. The independent results will take days to weeks to emerge.
