Llama 4 Runs on Arm
AI is moving faster—and getting smarter. Today’s open large language models are not only powerful but also designed with real-world deployment in mind: they’re lightweight, cost-efficient, and built to scale across billions of devices. In short, they’re ready for just about anything developers can imagine.
The launch of Meta’s Llama 4 is a great example—especially when you consider what it can do on Arm-powered platforms. With its innovative Mixture of Experts (MoE) architecture, Llama 4 delivers impressive performance in areas like multimodal reasoning, tool use, and more. But what really makes it stand out is how easily it can be deployed in real-world scenarios—thanks in large part to Arm.
Optimized for performance, ready for deployment
Arm’s flexible, power-efficient compute platform enables Llama 4 to run efficiently on Arm-based cloud infrastructure, giving developers the ability to deploy large language models with enhanced performance, lower power usage, and greater scalability across diverse cloud environments.
At a broader level, we’re seeing an interesting shift in the industry. While the push towards larger and more intelligent multimodal models continues, Llama 4 represents a rising trend toward smaller, practical models that enterprises and customers can run on their own infrastructure—whether in the cloud or on-premises. Llama 4, particularly the Scout model, is efficient, focused, and structured around agentic and MoE architectures that are exceptionally well-aligned with cost-efficient, scalable platforms like Arm.
Since the release of Llama 2, Arm’s dedication to optimizing model compatibility across its platforms ensures that developers and end users can efficiently deploy each new generation of Meta’s Llama models. Llama 4 Scout is a clear example of these optimizations in action, running seamlessly across the Arm ecosystem.
Llama 4 Scout is a milestone for Arm-based systems
We’re proud to announce that Llama 4 Scout runs efficiently on Arm-based infrastructure. To validate this compatibility, we successfully deployed Llama 4 Scout on Arm-based Graviton4 using the open source inference engine llama.cpp. This straightforward deployment demonstrates that developers can seamlessly integrate advanced AI capabilities without needing specialized hardware or proprietary software. With vertically integrated frameworks like llama.cpp and general-purpose ML tools like PyTorch, the path to production is clear and accessible.
Why Mixture of Experts (MoE) architecture is ideal for the Arm platform
- Intelligent Efficiency: MoE models intelligently route inputs to specialized subnetworks, dynamically allocating computational resources. This adaptive approach naturally complements Arm’s renowned energy efficiency and resource-conscious workload management.
- Scalable by Design: Arm platforms, such as AWS Graviton, Google Axion and Microsoft Cobalt, offer scalable core counts and threading capabilities ideal for the parallel nature of MoE models, effectively managing workloads to maximize both throughput and overall efficiency.
- Optimized for Diverse Workloads: Arm’s architecture philosophy emphasizes performance and efficiency across varied applications, closely aligning with the MoE capability to compartmentalize and specialize tasks within subnetworks.
- Forward-Looking Alignment: Together, Arm platforms and MoE architectures represent a forward-thinking synergy, equipped to meet evolving demands for smarter, more resource-efficient AI solutions.
Explore Llama 4 on Arm
We’re excited to invite developers and ecosystem partners to explore Llama 4 Scout on Arm—a powerful example of our shared commitment to open, collaborative AI. Running on Arm-based infrastructure like AWS Graviton, Llama 4 Scout delivers the performance, efficiency, and scalability needed for modern AI workloads.
Discover the expansive potential of Arm-powered AI and help shape a smarter, more connected future—from cloud deployments all the way to the edge.
Ready to get started? Explore the tools, connect with the community, and help shape a smarter, more connected future—powered by Arm.
Any re-use permitted for informational and non-commercial or personal use only.
Editorial Contact
Latest on X
Smaller models. Bigger impact. 💥
@Arcee_AI is using Arm-based CPUs to run SLMs in parallel for agentic AI. With 4x performance gains from quantization and KleidiAI, @julsimon shares why the future of AI is efficient, scalable - and built on Arm CPUs: https://okt.to/UtMRxA
There's just over a week left to enter the Arm Silicon Startups Contest! 🚨
If you're an early-stage startup, this is your chance to bring your silicon solution to market quicker, with $250,000 in Arm technology credit available to the winner.
Details: https://okt.to/FBoaGD
At 19, most of us were still figuring life out. @alexandr_wang was already building @Scale_AI.
He joins Rene Haas on the Tech Unheard podcast to talk all things AI, leadership and why being young is often one of his biggest strengths.
🎧 Listen now: https://okt.to/a7dY2B
What happens in Vegas… deserves to be shared!
At #GoogleCloudNext, the team celebrated 1 year of Arm Neoverse-based Google Axion processors with updates on migration and momentum, as more companies turn to Arm to balance performance, efficiency and cost:
It's full speed ahead for Windows on Arm. ⏩
With the new Windows Arm64 hosted runners for @github Actions, developers have the tools to easily build and deploy for Arm, on Arm - including powerful gen AI applications as the edge. Get the details in the blog below.
🚨 Your CI/CD workflows just got an upgrade!
Windows Arm64 hosted runners are now available for @GitHub Actions, giving you access to a familiar CI/CD pipeline to easily add Arm-native support and boost performance.
Here's to easier innovation on Arm: https://okt.to/wWXs5f
The new Arm-based @MediaTek Kompanio Ultra SoC is set to enable powerful generative AI experiences on Chromebook Plus devices.
Chris Bergey and Adam King share how the chipset unlocks next-level performance, efficiency, and new AI-enabled experiences: https://okt.to/gSywrF
As the automotive industry shifts to an AI-first future, Dipti Vachani, SVP & GM for Automotive at Arm, shares how we’re powering the transformation.
Speaking to @EETimes, Dipti explains how electrification and in-vehicle features rely on our technology.
Arm Powers Software-Defined Vehicle Revolution
Arm enables the automotive industry's fundamental shift towards software-defined and AI-driven vehicle with key strategies and technologies.
www.eetimes.com
In just a year since launch, Arm-based Google Axion CPUs have delivered substantial improvements in performance, efficiency, and scalability - just ask @Databricks, @Spotify or @ParamountPics.
We've put it to the test & the results speak for themselves: https://okt.to/P9mZuJ
We’re proud to join the UK’s AI Energy Council, working with other tech leaders to support the future of AI.
As the industry’s most power-efficient, high-performing compute platform, we're excited to bring our expertise to the Council and shape the future of computing.
Power drives AI’s success🔌
That’s why we’ve launched the AI Energy Council - a group of experts that will advise government on how to sustainably power AI and make use of the energy infrastructure we already have.
AI is pushing computational limits, demanding efficient, scalable architectures.
Get insights from @BenBajarin in the Silicon Reimagined report and discover how Arm CPUs drive AI inference and acceleration, powering the next generation of reasoning models:…
What would you say to your younger self? 💬
@Zoox CEO @Aicha2Evans best advice? Embrace the intensity, but know when it's time to relax and hit pause.
Catch the inspiring, real and insight packed conversation here - or wherever you get your podcasts! 🎧 https://okt.to/8t2grG
Congratulations, @AIatMeta on the launch of Llama 4! 👏
This is an exciting step forward in AI model design, and we look forward to our continued collaboration to optimize these new models across the Arm ecosystem, from cloud to edge.
By combining the flexibility, efficiency,…
Arm-Neoverse based Google Axion processors deliver the performance and efficiency needed to power your RAG applications with enhanced AI inference on LLMs.
📍 Catch us at #GoogleCloudNext later this week to find out more about making AI work smarter for you with Arm Neoverse.
RAG brings real-world data into AI - fast.
We put the framework to the test on Arm Neoverse-based Google Axion Processors, testing performance and cost efficiency. The result? 2.5x higher performance and 64% cost savings compared to x86: https://okt.to/w9gpMF
⭐ Let’s-a go! Nintendo Switch 2, unveiled earlier this week, delivers a serious power-up in performance and visuals, marking a new era in gaming. No mushroom required. 🍄
Congrats, @Nintendo, on another iconic leap forward.
Nintendo Switch 2 launches June 5 at $449.99, bringing new forms of game communication to life -...
System Provides Powerful New Hardware, Reimagined Online Capabilities and a Broad Range of Games All Together, Anytime...
okt.to
1. Robust infrastructure
2. Scalable compute
3. Power-efficient solutions
The AI Readiness Index found that 82% of leaders use AI, but only 39% have a strategy. Kevork Kechichian explains how to unlock AI's full potential with the right tech foundation: https://okt.to/iMVB1h
We’re working with @AstonMartinF1 to optimize performance on the track, in the pit, and in every critical decision.🏁
Charlie Blackwell shares why half a second saved can be crucial on race weekend - & how the Arm compute platform helps make that possible:…
We had a great time hosting @imec_int's Automotive Chiplet Forum at our Cambridge Office this week!
Chiplets will play a fundamental role in tackling the rising compute demands of our cars. We're excited to be driving this forward as part of imec's Automotive Chiplet Programme…
🤝 The Arm-based Kompanio Ultra from @MediaTek brings Armv9 CPUs and the Immortalis-G925 GPU to Chromebooks for the first time
Get ready to unlock enhanced generative AI experiences and next-generation performance on Chromebook Plus devices: https://okt.to/9ucfAq
In the age of AI, power efficiency isn’t just a competitive advantage. It’s a baseline requirement.
🆕 That’s why close to 50% of the compute shipped to top hyperscalers in 2025 will be Arm-based.
This is what reshaping the cloud landscape looks like: https://okt.to/sVuXFC
We’re rolling out new resources to make migration to Arm-based cloud platforms easier than ever. 💪
With step-by-step Learning Paths and acess to Arm experts, the time to start is now. Dive into the blog below to see how, and stay tuned for more updates coming soon!
Unlock performance and efficiency with our 🆕 cloud migration resources!
The new developer initiative includes:
🧠 100+ Learning Paths in our Cloud Migration Resource Hub
💬 Community support from Arm experts
🛠️ Access to engineering experts
https://okt.to/fk3INA
AI workloads are in need of a power efficient, sustainable solution.
The path forward lies in balancing power, performance, and economic viability to innovate and make a lasting impact. ⚡
Learn more: https://okt.to/l4EDpG
AI is a transformative force reshaping how we work, communicate, and solve problems.
Our AI Readiness Index offers actionable insights to help businesses adapt, innovate, and lead in the new world of AI.
Explore the findings ➡️
For 40 years, silicon innovation has redefined computing. Today, AI is driving a new era of custom silicon.
From mobile chipsets to next gen AI-optimized silicon designs, EVP of Solutions Engineering Kevork Kechichian breaks down how Arm is shaping the future 💬…
What does it take to become a CEO?
On the latest Tech Unheard episode, @aicha2evans joins Rene Haas to share her incredible path to CEO of @Zoox. Along the way, the two find powerful parallels in their journeys through the world of tech.
🎧 Listen now: https://okt.to/l1ic8U
