AMD Releases ROCm 10 and Makes ROCm.AI Generally Available

AMD ships ROCm 10 with ROCm.AI generally available

AMD released ROCm 10 on August 27, 2026 and made ROCm.AI generally available.

ROCm.AI is AMD's AI-native developer layer for building, running and optimizing AI workloads on AMD hardware. It combines three major components: ROCm Hyperloom, AMD Skills and the ROCm CLI.

There is an important availability distinction inside that bundle: ROCm.AI is generally available, while the ROCm CLI is currently a Technology Preview.

Hyperloom brings agentic optimization to inference

AMD describes ROCm Hyperloom as an autonomous agentic system for end-to-end inference optimization. It profiles a workload, finds bottlenecks, explores changes, implements targeted optimizations, benchmarks them and validates performance and correctness.

With ROCm 10, AMD says Hyperloom expands across AMD Instinct GPUs and supports vLLM and SGLang workflows, with optimization targets spanning HIP, Triton and FlyDSL.

The system can be used through standalone workflows or AMD Skills.

AMD Skills bring ROCm knowledge into coding agents

AMD Skills packages curated AMD-specific guidance and validated workflows for AI coding agents. AMD says the skills are now available through supported Claude Code, Codex and Cursor marketplaces as well as an open GitHub catalog.

The catalog covers client-native AI integration, cross-stack diagnostics and optimization, and server workflows for AMD Instinct GPUs and EPYC processors.

ROCm CLI is still a Technology Preview

The ROCm CLI provides one interface for inspecting systems, managing ROCm environments, serving models, running diagnostics and controlling runtimes.

AMD says the prebuilt CLI supports Windows and Linux and can manage side-by-side ROCm runtimes. However, the company labels the CLI Technology Preview and says its version-agnostic experience begins with ROCm 7.13, with official ROCm 10 support coming soon.

That means developers should not confuse the general availability of ROCm.AI with general availability of every component.

AMD reports large performance gains, with important test constraints

AMD reports that a system configured with ROCm.AI delivered an average 3.3x inference improvement and 2.4x training improvement over ROCm 7 on the same hardware.

These are AMD-reported laboratory results across specific MI355X configurations and selected models, not a universal speedup for every ROCm workload.

For inference, AMD's footnote describes an eight-GPU MI355X system and measurements across GLM-5, Kimi-K2.5 and DeepSeek-R1-0528. The training figure is based on a separate MI355X setup using Megatron-LM and selected DeepSeek and Qwen models. AMD notes that server configuration, software versions, drivers and optimizations can change results.

Why ROCm 10 matters

ROCm 10 also advances the broader software stack with a more modular ROCm Core SDK plus updates to libraries, compilers, frameworks, model support, performance tooling and hardware coverage.

The strategic shift is the addition of agentic optimization and coding-agent workflows directly into AMD's software experience. Rather than requiring developers to manually diagnose every GPU bottleneck, AMD is increasingly packaging platform expertise into tools that can inspect, propose and test optimizations.

For teams evaluating AMD accelerators, ROCm 10 is therefore both a conventional platform release and a test of a more agent-assisted GPU software workflow.