AI・テクノロジー
Meta's "Llama 4 Maverick" Full Release — Open-Weight LLM Ushers in the Era of In-House Inference
機械翻訳 / Machine-translated
Meta released the full weights of "Maverick," the top-tier model in the Llama 4 series, on September 11, 2026. With an estimated 400 billion parameters and commercial licensing included, this configuration has, for the first time, brought the prospect of "running GPT-5-class models on your own servers" within the realm of realistic cost calculations. The question of "buy API access or run in-house inference?" has now become something organizations can seriously debate at a practical level.
At 10:00 PM Japan time on September 11, Meta released the full weights of Maverick on Hugging Face. The accompanying technical report records MMLU at 89.3% and GSM8K at 96.8%. The gap with GPT-5 on average across major benchmarks is within 2.1 points — effectively overturning the industry consensus as of late 2025 that "there remains a gap in frontier performance."
"I ran Maverick on 8× RTX 6090s. 23 tokens per second. Roughly one-eighth the monthly equivalent of API costs. This is worth considering as an organization-wide migration." — AI infrastructure engineer at a major SaaS company (X, 22K followers)
The license is a proprietary "Llama 4 Community License," requiring prior application to Meta only for operators with more than 700 million monthly active users. In practice, this applies only to a handful of companies such as Google, Microsoft, Amazon, and ByteDance — for the vast majority of businesses, it poses no meaningful restriction.
Since the release of Llama 3 (2024), the performance curve of open-weight LLMs has been steep. Through the first half of 2026, Meta incrementally released Scout (70B-class) and Ranger (200B-class), with Maverick serving as the final chapter of that arc.
On inference costs, while GPT-5 API input pricing is estimated at around $5/MTok, multiple engineers have reported estimates of $0.6–$1.2/MTok for in-house inference of Maverick with INT4 quantization applied. Using AWS Spot Instances (p5.48xlarge), operation at approximately $32 per hour is also calculable on paper.
The official specifications confirm a maximum context window of 512K tokens. For agent workflows requiring large-scale codebase analysis or long-document processing, the range of what can be handled in a single prompt expands significantly.
While 8 or more GPUs are recommended for full-precision inference, reports indicate that 15–20 tokens per second are achievable on a 4-GPU setup with INT4 quantization applied. Mid-sized companies with on-premises GPU servers are beginning to see this as a viable option.
The LoRA and QLoRA ecosystem built up since Llama 3 supports Maverick from day one, with industry-specific fine-tuning code appearing on GitHub within hours of release. A rapid proliferation of derivative models is expected over the coming weeks.
In finance, healthcare, and public administration — sectors where regulatory constraints make sending data to external APIs difficult — the ability to achieve GPT-5-class performance within one's own environment carries significant weight. In Japan as well, in-house inference needs exist from a Personal Information Protection Act perspective, and discussions about adoption are expected to move to a practical level.
The significance of this release is not merely the "arrival of a powerful open model." It represents a structural shift in that it has elevated the question — "how long will we continue designing around the assumption of proprietary APIs?" — to the level of organizational decision-making.
The inference optimizations offered by OpenAI and Anthropic — the chain-of-thought reasoning of the o-series, Claude's long-context reliability — still function as differentiators. However, for broad use cases such as general-purpose text generation, summarization, and information extraction, open-weight choices will likely transition to "the self-evident default" over the next 6–12 months.
Of particular note are mid-sized SaaS companies and startups with high cost sensitivity. For operators whose monthly API expenses have reached the scale of several million yen, today marks the point at which the ROI of migration — even factoring in infrastructure investment — begins to make sense.
With the arrival of Llama 4 Maverick, the central question of AI procurement from 2027 onward will no longer be "which model to use" but "which inference infrastructure to build on." Organizations with GPUs should start recalculating costs immediately; those without face pressure to revisit the roadmaps of their cloud inference services. Will your organization continue to depend on APIs, or will it take the step toward in-house inference?
This article was written by an AI writer (AI News) from the Mirai News editorial team.
