Moore Threads unveils next-generation GPU architecture in first post-IPO app...

by XU Meihui

China-based GPU developer Moore Threads made its first major public appearance since listing, outlining a full-stack roadmap that underscores how Chinese chipmakers are trying to close the gap with global leaders through software, systems engineering and scale.

At its inaugural MUSA Developer Conference (MDC 2025) in Beijing on Dec 20–21, founder and CEO ZHANG Jianzhong presented a strategy spanning chip architecture, computing clusters and developer tools, all built around the company's proprietary MUSA platform.

The centrepiece was Huagang, Moore Threads' next-generation GPU architecture. Huagang supports full-precision computing from FP4 to FP64, lifts compute density by 50%, improves energy efficiency by a factor of ten, and is designed to scale to intelligent-computing clusters with more than 100,000 GPUs.

The architecture will underpin two planned products: Huashan, targeting integrated AI training and inference workloads, and Lushan, focused on high-performance graphics rendering.

Photo from Moore Treads

Moore Threads also released performance data from its KUAE 2.0 computing cluster at a ten-thousand-GPU scale. Zhang said training efficiency reached 60% on dense large models and 40% on mixture-of-experts models, with effective training time exceeding 90% and linear scaling efficiency approaching 95% — figures the company says are meant to demonstrate viability beyond laboratory benchmarks.

On the inference side, Moore Threads said it worked with AI infrastructure firm SiliconFlow to run the DeepSeek R1 671B model on its MTT S5000 accelerator, reporting per-card prefill throughput of more than 4,000 tokens per second and decode throughput above 1,000 tokens per second. The test was intended to signal readiness for large-scale inference workloads.

Zhang also outlined plans for an MTT C256 super-node architecture for next-generation hyperscale data centres, while flagging longer-term bets in embodied intelligence, AI for science (AI4S) and AI for 6G.

Beyond data centres, Moore Threads introduced the MTT AIBOOK, an AI-focused laptop for developers. Powered by its self-developed Changjiang system-on-chip, the device delivers 50 TOPS of heterogeneous AI compute and runs a Linux-based MT AIOS operating system with a pre-installed AI development toolchain. It is available for pre-order on JD.com at 9,999 yuan for a 32GB, 1TB configuration.

A Moore Threads R&D engineer told Jiemian News the device is intended primarily as an ecosystem validation tool, bundling hardware, drivers and software into a single machine to lower adoption risk for partners.

Photo by XU Meihui/Jiemian News

Software remains the hardest challenge. Nvidia's dominance rests not only on hardware but on CUDA, a platform built over more than a decade with roughly five million developers worldwide.

Moore Threads is positioning MUSA, now upgraded to version 5.0, as a China-based full-stack alternative spanning architecture, instruction sets, runtime libraries and drivers. Zhang said ecosystems form the GPU industry's core moat, with long-term value increasingly driven by software and developer adoption rather than hardware alone. He added that efficiency in key AI operations has improved and core components will be gradually open-sourced.

Industry data underline the stakes. TrendForce expects global AI server shipments to maintain double-digit growth in 2026, while IDC estimates China's accelerated server market reached about US$16 billion in the first half of 2025.

Partners interviewed by Jiemian News were candid about the gap. One said Chinese GPUs still trail leading global products by roughly one to two generations in performance and ecosystem maturity, but are already considered cost-effective for many real-world inference workloads.

After its market debut, Moore Threads now faces a tougher test: turning ambitious roadmaps into reliable delivery, and scaling an ecosystem under pressure from entrenched global leaders and intensifying domestic competition.