AA-AgentPerf adds AMD MI355X benchmark results for DeepSeek V4.1 Flash

AMD: AMD is a semiconductor company focused on developing high-performance computing hardware, including its Instinct series of AI accelerators based on the CDNA architecture. In the context of this news, AMD is debuting benchmark results for its latest MI355X accelerator on agentic inference workloads. Recent developments highlight ongoing software optimizations through ROCm and community contributions to frameworks like vLLM and SGLang. MI355X: MI355X is AMD's flagship AI accelerator in the MI350 series, built on the CDNA 4 architecture and positioned as a competitor in the high-end inference market. It is the subject of new performance evaluations in this news, with results added to the AA-AgentPerf leaderboard for multiple large models. Recent industry activity includes targeted optimizations for agentic workloads on this hardware. GLM-5.2: GLM-5.2 is an advanced AI model from the GLM series developed by Zhipu AI, used in coding and agentic workloads. The news includes performance comparisons for this model on AMD MI355X hardware versus NVIDIA systems. It appears in recent rankings of leading Chinese AI models. GLM-5.3: GLM-5.3 is a later iteration in the GLM series from Zhipu AI, optimized for high performance in complex inference scenarios. New results on the AA-AgentPerf benchmark for this model on MI355X are featured in the news. It is frequently highlighted in October 2026 assessments of top Chinese AI models. Kimi K3: Kimi K3 is a large-scale AI model developed by Moonshot AI, noted for its extensive parameter count and capabilities in long-context tasks. It is one of the featured models in the new MI355X benchmark results reported in this news. Recent evaluations position it among top-performing Chinese open-weight models. AA-AgentPerf: AA-AgentPerf is an agentic inference performance benchmark developed by Artificial Analysis that evaluates hardware and software using realistic coding agent trajectories. This news expands the benchmark's coverage by including AMD MI355X results alongside NVIDIA systems for models like Kimi K3 and GLM variants. The benchmark emphasizes real-world factors such as prompt caching and tool-call delays. DeepSeek V4.1 Flash: DeepSeek V4.1 Flash is a high-speed AI model developed by DeepSeek, designed for efficient inference in demanding applications. The news adds new benchmark results for this model on AMD's MI355X hardware as part of ongoing AA-AgentPerf evaluations. It ranks among leading Chinese AI models in recent independent assessments. Model Ecosystem: Leading Chinese AI models including Kimi K3 and GLM series variants are undergoing cross-vendor hardware evaluations to assess efficiency in agentic applications. Benchmark Evolution: Specialized benchmarks focused on realistic agentic workloads are gaining traction as the industry shifts toward evaluating systems on long-context, multi-turn tasks. Hardware Competition: AMD continues to advance software stacks such as ROCm and integrations with serving frameworks to improve AI inference performance on its accelerators.