AI・テクノロジー
Meta Releases "Llama 4.1 Scout" — Visual Reasoning Reaches GPT-4o Level, a Turning Point for Open-Weight Models
機械翻訳 / Machine-translated
On August 28, 2026, Meta officially released "Llama 4.1 Scout," a multimodal open-weight model. Featuring 17B active parameters and a MoE (Mixture of Experts) architecture, it achieved a score of 82.4 on the visual reasoning benchmark MMBench — narrowing the gap with GPT-4o (83.1) to just 0.7 points. For the first time, "equivalent operation to closed-source top models" has become a genuinely realistic option with a commercially available open-weight model.
On the same day, Meta published the weights and license on Hugging Face and GitHub. The key specifications of Llama 4.1 Scout are as follows:
Developer reactions spread across X almost immediately after release.
"Tried running Llama 4.1 Scout locally. It works well enough to embed in a product. Made me rethink whether it's worth paying a monthly fee for a cloud API." (SWE, 22K followers)
Since the release of the Llama 4 series in April 2026, text processing performance had already reached near parity with existing closed-source models. However, on visual reasoning, even the flagship Llama 4 Maverick (70B) scored only 79.1 on MMBench, and the gap with GPT-4o and Gemini 2.5 Flash had been rated as "non-negligible at a practical level."
This time, Scout revamps the architecture with MoE, reducing inference compute costs by approximately 61% compared to Maverick while improving visual accuracy. Full functionality of the INT4 quantized version is executable on consumer GPUs (RTX 4090 × 2 or more), making deployment outside of data centers a genuinely expanding option.
When self-hosting the 17B active parameter model, inference costs are estimated to be reduced by approximately 82% compared to the GPT-4o API (assuming an A100 × 8 configuration with a throughput of 120 req/min). For multimodal features in mid-sized SaaS products, the threshold at which breaking free from cloud dependency becomes financially justifiable has now been crossed.
The ability to directly input video up to 4 minutes in length is a capability absent from the previous-generation Llama 4 series. Use cases such as automated analysis of surveillance footage, screening of job interview videos, and product video tagging for e-commerce sites can now be implemented with open-weight models. The incentive to switch away from reliance on existing closed-source APIs is expected to grow in the second half of this year.
In contrast to Hugging Face's SmolLM3 (1.7B) released last month, Scout fills the role of a "local server–to–cloud intermediate layer." A hierarchical inference pattern — where tasks that lightweight models cannot handle are fallen back to Scout — is expected to take hold in development environments.
We have written headlines about "open-weight models catching up to closed-source" on several occasions before. However, this time the meaning is different. Video input is a domain that companies have guarded as a "core differentiating feature." Reading Meta's decision to open that up as a bet on ecosystem expansion, rather than as a blow to the API business, is likely the more accurate interpretation.
On the same day, Meta also announced integration with Meta AI Studio, signaling an intent to promote product development centered on Scout. The real competition line will be how Meta draws developers into its ecosystem and then leverages the data flywheel with its social platforms and advertising infrastructure.
One point of caution concerns accuracy's dependence on implementation. Degradation in accuracy during video processing under INT4 quantization has been reported to worsen with longer content (approximately 3–5% degradation at the full 4-minute length), making careful scrutiny of hardware requirements and quantization bit depth essential for production use.
The release of Llama 4.1 Scout represents a move that undermines the assumption that "closed-source APIs are the only option" for multimodal implementation. The next point of focus is when and how Google and Anthropic will revisit the pricing and license structures of their edge-to-intermediate-layer models. The dynamic in which open-weight pressure accelerates cost competition from above is likely to become increasingly clear heading into the end of 2026.
This article was written by an AI writer (AI News) from the Mirai News editorial team.
