The day Meta declared war on the AI giants
Meta unveiled on Monday, October 12, Muse Wonder and Muse Wonder Full, its two most capable models ever produced and its first explicit foray into the frontier artificial intelligence market. Internally nicknamed "Watermelon" during 14 months of development, the two systems are described by the company as a leap toward "personal superintelligence" — the ambition to offer every user an AI assistant capable of remembering, reasoning, and acting over the course of entire weeks and months.
The announcement is not merely technical; it is strategic. Since the large language model market exploded in 2023 with ChatGPT, Meta has held the position of the open-source company, relying on models like Llama to compete more affordably while OpenAI and Anthropic accumulated advantages with proprietary next-generation systems. Now Meta wants to lead that race — at least when it comes to raw capability.
Active Memory: the long-term bet
The announced differentiator for Muse Wonder is a capability Meta calls "Active Memory." Unlike context mechanisms that maintain short conversations, Active Memory is designed to retain and synthesize information over days, weeks, and months. On the MemoryBench benchmark, created specifically to evaluate this continuous learning capacity, Muse Wonder outperformed both GPT-6 Astra and Claude Opus 5.5.
"We put a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months, and beyond," said research scientist Kavya Kopparapu in a launch video. The promise fits with Meta's long-term strategy around the Muse agent — an assistant that does not just answer questions but executes complex tasks: navigating insurance claims, analyzing lease agreements, building working applications.
Coding capability and benchmarks
In programming, Muse Wonder claims parity with ChatGPT and Claude on professional and coding tasks, according to chief AI officer Alexandr Wang. Meta's team tested the models in the context of the Prometheus data center, a gigawatt-scale facility designed specifically to train large models, and the results were comparable to the market's leading competitors.
Availability, however, follows a deliberate two-tier strategy. The standard Muse Wonder will be offered via API to developers and builders, priced at $4 per million input tokens and $20 per million output tokens — a tier equivalent to GPT-6 Sol and Opus 5.5. Muse Wonder Full, the more capable version, will be restricted to Muse Code subscribers and will not be made available publicly via API.
This duality reflects a business calculation: use the basic Wonder to attract a developer ecosystem and keep the Full as an exclusive competitive advantage for Meta's own Muse platform.
The weight mystery and European regulation
A crucial point remains unanswered: Meta has not announced whether it will publish the model weights. That decision has direct implications under the European Union's AI Act. Truly free and open models may escape some of the documentation requirements imposed by Article 53 on the EU AI Office — something Meta has not yet decided for Wonder, nor for Muse Spark 1.3.
Meta's path through open-source has been intermittent. The company released a distilled 30-billion-parameter model but kept the weights closed for larger models. Zuckerberg has signaled caution about opening the most powerful systems, especially as they approach capabilities the company itself classifies as "frontier."
Safety at the frontier
The announcement included a dedicated section on safety. Meta claims it has trained Muse Wonder to be the most aligned model it has ever produced, with robust protections against malicious use of chemical, biological, radiological, and nuclear (CBRN) capabilities and against cyber exploitation. The company's policy still preserves access to frontier cyber capabilities for defensive use, per the Meta Superintelligence Scaling Framework.
The safety context is especially relevant given recent incidents involving autonomous AI agents — from both Anthropic and OpenAI — that behaved unexpectedly when given open-ended instructions. Meta has not disclosed whether it observed Muse Wonder-powered agents behaving anomalously during internal testing, nor detailed the specific measures taken. Zuckerberg announced last month that it had delayed the release of the Muse models to invest in safety.
What to watch
With the launch of Muse Wonder, Meta places itself officially in the frontier model arena — a space dominated by OpenAI and Anthropic until now. The question is no longer whether the company will compete, but how long it will take for Wonder to gain traction in a market that has already consecrated the incumbents. The answer hinges on at least three factors: whether the weights will be open, whether Active Memory truly delivers what it promises in real-world scenarios, and whether the two-tier strategy (open API vs. closed Muse Code) will build the developer base necessary to sustain the platform.
The museum of personal superintelligence is now open — but we still do not know who will live inside it.
Sources: TechCrunch, WIRED, The Next Web
✓ Independent sources cross-checked and verified before publishing
