AI・テクノロジー
Meta "Llama 4 Maverick 2" — All Commercial Restrictions Lifted, Open-Weight Model Reaches Top Inference Performance
機械翻訳 / Machine-translated
Meta released "Llama 4 Maverick 2," the latest in the Llama 4 series, on September 3, 2026. The commercial licensing requirement previously imposed on operators with more than 700 million monthly active users has been completely abolished. The model became available for immediate download on HuggingFace, and it is believed that over 100,000 downloads were recorded within three hours of release. The shift from "buying an API" to "owning a model" has now come firmly within reach.
Meta announced the release via its official blog at 18:00 JST on September 3, 2026. Key specifications are as follows.
Reports of real-world verification appeared immediately on X.
Deployed Llama 4 Maverick 2 on our own servers. Answer quality is nearly on par with GPT-5 mini, while estimated inference costs are less than 1/9th. This changes our decision on whether to integrate it into our SaaS.
The download pace is believed to be more than double that of the Llama 3.3 release, with the numbers reflecting strong interest from enterprise users.
Since Llama 2, Meta had required large-scale operators with more than 700 million monthly active users to obtain individual commercial licenses. Microsoft, Amazon, and Google each entered into separate agreements, which effectively functioned as a barrier to entry.
The removal of these restrictions can be read as Meta's move to reclaim leadership of the ecosystem using "completely free" as its weapon, amid Mistral Large 3 (July 2026) and DeepSeek R3 (August 2026) building market share through an open approach. The competitive axis is shifting from technological competition to institutional competition.
MATH 91.2 exceeds OpenAI's GPT-5 mini (89.8) by 1.4 points, and HumanEval 87.3 reaches the level of practical coding competence. The model is expected to be immediately deployable across three domains: legal summarization, mathematical analysis, and code generation.
Inference costs via self-hosting are estimated to be 1/8th to 1/10th compared to API-based access. For operators handling around 100 million queries per month, the annual cost difference could amount to tens of millions of yen. This is a pivotal moment that fundamentally transforms the cost structure of AI embedded in SaaS.
With the elimination of commercial restrictions, commercially deploying industry-specific fine-tuned models based on Maverick 2 is now officially possible for the first time. Vertically specialized models in the medical, legal, and manufacturing sectors are expected to emerge in large numbers over the next three to six months.
The licensing burden of the Llama series used as a base model by major domestic companies such as NTT, Fujitsu, and CyberAgent disappears. The mass production cost of Japanese-language fine-tuning is set to decrease, and competition among domestic LLMs is expected to accelerate.
If you distill the essence of this announcement, the structural change that matters more is not "performance has caught up with GPT-5 mini," but rather "the license is gone." The fact that prompts are not sent to external API servers becomes a decisive differentiating factor in terms of data sovereignty for the financial, medical, and legal sectors. The era has arrived in which organizations that were previously forced to choose closed APIs for reasons of regulatory compliance and confidentiality protection can now choose self-hosting on both performance and economic grounds.
However, caveats apply. Because Maverick 2 uses a MoE architecture, deploying the full 400B requires securing multiple A100-class GPUs. The "true TCO" — inclusive of infrastructure costs and personnel expenses — may well exceed API costs for mid-sized or smaller organizations. The decision to "own a model" requires not only an assessment of procurement costs, but also an honest inventory of operational capabilities.
DeepSeek cut the path, Mistral widened it, and Meta broke the seal. There is no longer any room to doubt that this trajectory moves in one direction.
An unrestricted commercial open-weight model has drawn level with GPT-5 mini in inference performance. There is no shortage of organizations that now need to answer the question — "use an API or own a model?" — as a business decision, starting today. Has the premise of your AI procurement changed with today's announcement?
This article was written by an AI writer (AI News) from the Mirai News editorial team.
