Arm AI Portal Puts Optimized Models in Front of Agents

Arm on September 8 opened Arm AI Portal, a catalog of tuned models that coding agents can query through MCP. Sharbani Roy, VP of AI and Developer Platforms at Arm, said more than 22 million developers now need that software across cloud, edge, and physical AI.

The pages list Qwen, Gemma, and YOLO with latency, memory, and size. The same records are wired into Copilot, Codex, and Claude Code through a new MCP endpoint, so the agent that writes the app can pick the Arm build without opening a browser.

Arm Put Models, Numbers, and Agents in One Place

Roy’s September 8 Arm AI Portal announcement treats search, benchmarking, and hand tuning as the delay that kills an on-device build. Developers get task-specific builds with accuracy figures, sample code, and deploy workflows. Agents get the same records as machine-readable signals.

The live pre-optimized models with performance data cover language, speech, vision, and neural graphics. At launch the named families are Alibaba’s Qwen, Google’s Gemma, and Ultralytics YOLO, running on ExecuTorch, LiteRT, and ONNX Runtime. Partners named on day one are Alibaba, Raspberry Pi, and Ultralytics.

WHAT SHIPS ON DAY ONE

• Model families: Alibaba Qwen, Google Gemma, and Ultralytics YOLO, tagged by task rather than by a single chip.
• Runtimes: ExecuTorch from PyTorch, LiteRT from Google AI Edge, and ONNX Runtime.
• Hardware filters: Cloud CPU, Mobile CPU, Mobile GPU, Edge Linux, and Embedded NPU.
• Distribution: Arm-optimized models on Hugging Face, plus Raspberry Pi and vivo pages inside the portal.

Bring-your-own models, including private weights, are promised next. Roy wrote that the portal will soon add analysis and tuning for those weights on Arm. That path is an Early Access program, not a button on the public catalog.

An MCP Endpoint for Copilot, Codex, and Claude

Arm already taught coding agents how to move software onto its ISA. The public Arm MCP Server for x86 migration searches docs, scans C++, Python, Go, JavaScript, and Java with migrate-ease, checks container architectures, and scores assembly with LLVM-MCA. The AI Portal endpoint is a second feed, aimed at models rather than ports.

On the developer site the install line is codex mcp add arm-ai --url https://mcp.api.devplatform.arm.com/ai-portal. Once that server is attached, the agent can list tuned builds, compare options, and pull workflows for a target without leaving the editor. Arm lists Claude Code, Codex, GitHub Copilot, Qoder, TRAE, and CodeBuddy as supported environments. Agent-ready resources are early access, ahead of a general release.

Building an AI application shouldn’t start with weeks of searching, benchmarking and optimization. Developers need to find the right model and understand its performance, while agents need clear signals to discover the same models, tools and information.

Sharbani Roy, VP, AI and Developer Platforms, Arm

That is the product bet. If Copilot or Codex is the first reader of a new project, the hardware that shows up in its tool list is the hardware that ships. A human can still browse Hugging Face. An agent will not, unless someone puts Arm’s numbers on MCP.

Allow us to introduce you to Arm AI Portal. 👋

Helping 22M+ developers and their agents build the future of agentic AI on Arm.

How?

By bringing together pre-optimized AI models, performance insights, code examples and deployment workflows in one place.

This is the path from… pic.twitter.com/1YHbonbIB5

— Arm (@Arm) September 9, 2026

The production pick is rarely the model that wins a public leaderboard. Latency, peak memory, energy, privacy, and which runtime actually runs on the board decide the build, and those fields are what the portal asks an agent to compare.

How Fast Qwen3-TTS Runs on a vivo X300

Arm published two on-device figures with the launch, both single-thread, both on hardware it already sells into. They are vendor measurements, not a third-party suite.

ARM’S LAUNCH SPEED CLAIMS

The phone result rides Scalable Matrix Extension 2, Arm’s matrix instructions on the CPU. The Pi 5 result does not. Raspberry Pi 5 is a NEON board here, so the same “over 40%” line is a quant-and-runtime win, not an SME2 win. Mixing those two stories into one “Arm is 40% faster” headline would be wrong.

Qwen3-TTS is speech. YOLO26n is detection. The portal also advertises neural graphics, but the launch note does not attach a graphics fps number to a named title. Comparison views on the site plot latency against peak memory and package size, which is the trade an agent can read without a lab.

New Phone Silicon Still Needs a Findable Model

The same day, Chris Bergey, Arm’s executive vice president for Edge AI, introduced CSS for Mobile 2, a compute subsystem for phones that puts a new C2 CPU cluster beside a Mali G2-Ultra NX GPU. Bergey wrote that new hardware only matters when developers can reach it through tools they already use. AI Portal is that on-ramp for SME2-tuned models and for GPUs that now carry neural accelerators.

CSS FOR MOBILE 2, PER ARM

• C2-Ultra AI: Up to 1.7x higher AI performance than C1-Ultra.
• Single-thread: 15% higher, with up to 38% less power at the same performance.
• SME2 in the cluster: Two units, and a 70% speedup on the latest small language models.
• Mali G2-Ultra NX: Dedicated neural accelerators, up to 4x higher performance per watt on neural graphics, and up to 14% higher performance on existing games.

Those GPU “up to 4x” watts are not the Qwen3-TTS “over 4x” on the vivo X300. One is a graphics efficiency claim for new Mali silicon. The other is a speech-model speedup on a shipping phone. SME2 is already in leading Android and iOS handsets, with Alipay, Google AI Edge Gallery, OPPO, and vivo named in the software trail.

KleidiAI is the quieter layer under the portal. It is an open-source kernel library wired into PyTorch, ExecuTorch, LiteRT via XNNPACK, ONNX Runtime, MediaPipe, MNN, and llama.cpp, so a lot of the speed lands without a developer rewrite. The portal does not replace those kernels. It tells a person, or an agent, which packaged model to grab so those kernels have something to run.

Qualcomm’s Hub Already Lists 300-Plus Models

Arm is late to a format Qualcomm has been shipping for years. Qualcomm’s own developer pages describe more than 300 models, validated on Snapdragon, with on-device latency and memory, sample apps, and a GitHub Qualcomm AI Hub model collection. Runtimes there include Qualcomm AI Engine Direct, LiteRT, and ONNX, plus llama.cpp paths on Hexagon.

TWO VENDOR CATALOGS

Qualcomm wins on count and on chip-specific NPU bundles. Arm’s counter is width. The same portal is supposed to serve a cloud Neoverse CPU, a Pi on a robot, and a vivo phone, because Arm licenses the ISA rather than selling one SoC. If an agent’s job is “detect objects on a Pi 5,” Qualcomm Hub is the wrong shop. If the job is “run this on Snapdragon 8 Elite,” Arm’s three families look thin.

NVIDIA’s NIM catalogs still own the GPU server default, which this launch does not pretend to displace. The fight Arm picked is CPU and edge inference, where CUDA is not the runtime and the board may not have a large NPU.

The Waitlist for Custom Model Tuning

The public catalog is a start. The parts that make a portal sticky for a company that already trained its own weights are still gated.

WHAT WE KNOW

• Live now: The developer catalog, Hugging Face mirrors, and the Codex install line for the AI Portal MCP.
• Named partners: Alibaba, Raspberry Pi, Ultralytics, with vivo called out as a supported phone target.
• Vendor speed figures: Over 4x for Qwen3-TTS on a vivo X300, and over 40% for YOLO26n on that phone and on a Raspberry Pi 5.

WHAT IS UNCONFIRMED

• General MCP release: Arm has not set a date when agent access leaves early access.
• Bring-your-own timing: Tuning for proprietary weights is “soon,” through Early Access, with no public SLA.
• Independent checks: The 4x and 40% figures have not been repeated in a public third-party suite tied to this launch.
• Catalog depth: Arm has not published a head count of individual model builds behind the three family names.

Until custom-model tools are public, a team with private weights still does the old work: export, quantize, bench, and hope KleidiAI catches the matmuls. The agent can only recommend what Arm has already packaged. That is a small menu next to Qualcomm’s, and it will stay small until the waitlist opens.

The portal is live, the MCP feed answers from a URL, and the phone silicon it is meant to feed is CSS for Mobile 2. Qwen, Gemma, and YOLO are the bait. Whether Copilot starts answering “run it on Arm” depends on how fast that menu grows past three names.

Frequently Asked Questions

What Is Arm AI Portal?

It is Arm’s public catalog and agent feed for tuned AI software, hosted at developer.arm.com/ai. The comparison view plots latency against peak memory in megabytes and package size in megabytes, and it filters by Cloud CPU, Mobile CPU, Mobile GPU, Edge Linux, and Embedded NPU, so a build can be chosen by board class rather than by model name alone.

How Do Coding Agents Connect to Arm AI Portal?

Arm documents a Codex install, codex mcp add arm-ai --url https://mcp.api.devplatform.arm.com/ai-portal, and lists Claude Code, Codex, GitHub Copilot, Qoder, TRAE, and CodeBuddy as supported environments. That agent path is early access, separate from the older GitHub MCP that scans codebases for x86-to-Arm ports.

How Is Arm AI Portal Different From KleidiAI?

KleidiAI is the open-source kernel library that accelerates matmuls and related ops inside frameworks such as PyTorch, ExecuTorch, LiteRT, ONNX Runtime, and llama.cpp, and Arm’s Kleidi pages cite up to 18 times faster Llama inference, 15 times for Gemma 2 2B, and 2.2 times for Bert-Large after those kernels land. The portal does not ship those kernels. It is the catalog, the numbers, and the MCP interface that point a developer or an agent at a packaged model those kernels can run.

Does the Raspberry Pi 5 YOLO Figure Use SME2?

No. Arm’s launch note credits the Pi 5 YOLO26n gain to NEON with mixed FP16 and INT8 versus FP32. SME2 is the accelerator named for the vivo X300 figures, both Qwen3-TTS and YOLO26n. Pi 5 remains a NEON target in this dataset.

Which Devices Already Include SME2?

Arm’s SME2 pages list iPhone 16, iPhone 17, M4-based devices, and the vivo X300 series, with more devices due through 2026. CSS for Mobile 2 doubles SME2 units in the C2 cluster for the next phone generation, which is why the portal is being stood up before those chips reach retail.