AI・テクノロジー
Meta Officially Releases "Llama 4 Ultra" — MMLU 93.2% Shifts the Competitive Landscape for Open-Source Reasoning Models
機械翻訳 / Machine-translated
On August 22, 2026, Meta officially released its open-source large language model "Llama 4 Ultra" on Hugging Face and the company's developer portal. With 405 billion total parameters (MoE architecture, approximately 86 billion effective active parameters) and an MMLU score of 92.3%, the option to "run GPT-4o-level reasoning on your own infrastructure without relying on closed APIs" is becoming a practical reality for organizations with sufficient infrastructure budgets.
At 11:00 PM Japan time on August 22, 2026 (that same morning, PT), Meta publicly released the weight files and model card for "Llama 4 Ultra" on Hugging Face. The license is the Llama 4 Community License (commercial use permitted, except for services with more than 700 million monthly active users).
Key specifications:
"We ran Llama 4 Ultra on 8×H200 at BF16 full precision. Speed is slightly slower than GPT-4o's API response, but quality is equivalent or better. The biggest difference is being able to fine-tune on internal data." (CTO of an ML startup, 12,000 followers)
Since the release of Llama 1 in February 2023, the Llama series has pushed the performance ceiling of open-source LLMs with each new version. The Llama 4 generation released in April 2025 (Scout / Maverick) was the first to significantly narrow the gap with closed models, and the Ultra is positioned as the largest configuration in that generation.
Meta's continued commitment to releasing models of this scale as open source is driven by strategic motivations: differentiating from AWS and Google Cloud in the cloud AI market, and integrating AI into its advertising and commerce infrastructure. "Democratizing AI" is not merely a slogan — it is widely seen as a move to shift the layer of ecosystem control downward.
Among open-source competitors, DeepSeek V4 (671 billion MoE) took an early lead in performance at the start of 2026, but Ultra differentiates itself through a combination of clear commercial licensing and inference efficiency.
An MMLU score of 93.2% reflects breadth in knowledge and reasoning, while a HumanEval score of 89.7% reflects depth in coding-specific tasks. While specialized models excelling at one or the other have existed, open models scoring around 90% on both have been virtually nonexistent — until now. This effectively marks the first time an open model has crossed the threshold to serve as a foundation model for general-purpose agents.
Despite 405 billion total parameters, only the equivalent of 86 billion are active during a single inference pass. Multiple ML engineers have reported estimates suggesting that running equivalent-quality tasks on an H200 ×8 setup could reduce costs by 60–70% annually compared to the GPT-4o API. However, infrastructure setup and maintenance costs are additional, so for small-to-medium scale operations, APIs remain a rational choice.
Full legal documents, entire codebases, and long-term conversation logs can all fit into a single request. However, actually utilizing context of this length involves tradeoffs between inference speed and memory requirements. In practice, "tens of thousands to hundreds of thousands of tokens" is currently considered the realistic operational upper limit.
The combination of a commercial license and publicly available weights means that companies can fine-tune on their own data and complete deployment entirely in-house. The use of confidential data — previously impossible with closed APIs — is expected to move into serious consideration in the fields of legal, compliance, and healthcare.
There is one reason to consider the release of Llama 4 Ultra a "turning point": the MMLU score of 93.2% has crossed the psychological threshold of "good enough to use."
Until now, open models have been treated as "slightly inferior alternatives to closed ones." Ultra has become the first open model to sit in a position of being "equal to or better than closed models" on benchmarks. It is expected that over the next six months, several large enterprises will make decisions to eliminate their dependence on GPT-4o API.
That said, caution is warranted. Between "being able to run it" and "being able to put it to practical use," lies the full scope of RAG pipeline design, safety filter tuning, and the establishment of monitoring systems. The decision to "add Ultra as an option" and the decision to "fully migrate to it" carry entirely different risk profiles.
How Meta moves next is also worth watching. Simultaneously with Ultra's release, a preview of "Llama Tune" — a fine-tuning optimization tool for Llama 4 — was quietly announced. Once this is ready, even non-ML engineers will have an entry point for in-house customization. This may be remembered as an inflection point at which the open-source ecosystem began to offer a level of convenience approaching that of closed managed services.
Llama 4 Ultra has brought "open-source, GPT-4o-class" performance into reality. Whether to continue using closed APIs or invest in infrastructure to bring open models in-house — the number of organizations facing that decision is expected to grow heading into the fall of 2026.
Is your organization prepared to choose which bet to make?
This article was written by an AI writer (AI News) from the Mirai News editorial team.
