Meta AI model updates: are the latest versions included in standard subscriptions in 2025?
- Graziano Stefanelli
- 20 hours ago
- 3 min read
Meta’s AI strategy has long combined open-source releases with platform integration, enabling everyday users to tap world-class language models across Facebook, Instagram, WhatsApp, and a standalone Meta AI app. In April 2025 the company shipped two new Llama 4 variants—Maverick and Scout—and immediately wired them into the free experience without introducing a paywall. Yet the most powerful sibling, Llama 4 Behemoth, remains in internal preview. This article clarifies exactly which models are available to standard (free) users, what remains behind closed doors, and how that policy might evolve.
Llama 4 Maverick and Scout are fully open and integrated into the free tier.
Meta released Llama 4 Maverick (34 B parameters) and Scout (7 B) under an open-source license on 5 April 2025. Within three weeks, Meta AI across all platforms flipped to Maverick as its default reasoning engine. Scout serves as a lightweight fallback on lower-end devices and edge deployments.
Model | Parameter count | Release status | Where you can use it | Context window |
Llama 4 Scout | 7 B | Open-source (April 2025) | Edge cache for smart glasses, web fallback | 128 K tokens |
Llama 4 Maverick | 34 B | Open-source (April 2025) | Default in Meta AI app, Messenger, Instagram, WhatsApp, Ray-Ban glasses | 1 M tokens |
Because both models carry an Apache 2.0 license, developers outside Meta can also fine-tune or embed them, making Maverick the de-facto “standard subscription” model—no payment or verification required.
Llama 4 Behemoth remains an internal preview reserved for research and red-team analysis.
Alongside its public release, Meta teased Llama 4 Behemoth, a >140 B-parameter giant intended to push multimodal reasoning and real-time planning. As of September 2025, Behemoth:
Has no public checkpoint—not even under a non-commercial license.
Powers selective internal pilots (e.g., Meta customer-service chatbots).
Runs safety and bias stress tests before any external rollout.
Meta’s roadmap states that Behemoth could appear in “controlled enterprise previews” sometime in 2026, but no consumer subscription—free or paid—currently exposes it.
Platform parity ensures the same latest model for all free users.
A highlight of 2025’s rollout is uniform model parity: whether you invoke Meta AI via the Messenger chatbot, the Instagram search bar, or the Ray-Ban smart-glasses voice assistant, you receive Maverick’s responses with identical context limits and multilingual support.
Platform surface | Underlying model | Paywall? | Unique extras |
Meta AI app (iOS/Android/web) | Maverick | No | Projects workspace |
Messenger / FB chat | Maverick | No | AI image suggestions |
WhatsApp assistant | Maverick | No | On-device caching |
Instagram search | Maverick | No | Reel caption rewrite |
Ray-Ban smart glasses | Scout → Maverick cloud escalate | No | Voice-first interface |
This stands in contrast to many competitors that gate the newest model (e.g., GPT-4o, Gemini 2.5 Pro) behind Pro tiers.
Why Meta still withholds Behemoth despite its open-source track record.
Mark Zuckerberg reaffirmed the company’s “open by default” stance, yet signalled caution for extra-large models due to safety and geopolitical considerations. In a July 2025 interview he noted three gating factors:
Weaponizable misuse. Larger context windows can accelerate disinformation or bio-risk synthesis.
Compute inequality. Open-sourcing giant checkpoints could benefit state-level actors beyond Meta’s governance reach.
Regulatory uncertainty. EU AI Act compliance for frontier-scale LLMs remains unresolved.
Thus Meta is piloting a “release by tiers” approach—smaller models open, mid-range models integrated free, frontier models withheld pending safeguards.
What standard users should expect through 2026.
Quarterly refresh cadence. Meta promises to swap in any public successor (e.g., Llama 4.1) across all surfaces without a subscription fee.
Incremental context boosts. Maverick already supports 1 M tokens; upcoming patch aims for 2 M via dynamic windowing.
Multimodal upgrades. Audio-and-vision fused models derived from Maverick will replace separate whisper-style pipelines by early 2026.
Unless Meta reverses policy, the company will continue delivering its latest open-source model to the free user base, keeping Meta AI one of the most accessible cutting-edge assistants.
Key takeaways for users considering Meta AI versus paid competitors.
Criteria | Meta AI (Free) | Paid competitors (GPT-4o Pro, Gemini Pro, Copilot Pro) |
Latest public model | Always included day-one (Maverick) | Paywall or limited queries |
Context window | Up to 1 M tokens | 128 K to 1 M, usually behind subscription |
Cost | Free | $20–$30 per user per month |
Model openness | Apache/MIT checkpoints | Closed weights |
Frontier access (Behemoth / Ultra) | Not yet available | Available in paid tiers |
For students, casual users, and developers who value transparency and zero cost, Meta AI offers unparalleled access to a top-tier LLM. Enterprises needing extra-large models or specialized compliance may still look toward paid ecosystems, but Meta’s free tier remains uniquely generous in 2025.
____________
FOLLOW US FOR MORE.
DATA STUDIOS