XPU AI accelerators
As of 2026-09-11Kunlunxin · AI accelerator portfolio (source)Company dossier · Reviewed 2026-10-31
Kunlunxin 昆仑芯
AI accelerators with a cloud and application ecosystem around Baidu.
What the evidence says
IDC-reported market share and inspectable RLinf support strengthen adoption and software evidence. The disclosures do not identify product mix, customer utilization, matched performance, power efficiency or broad software-ecosystem parity.
Reuters · Nvidia's China AI-chip share falls as Huawei closes gap, report finds (source)Kunlunxin · Kunlunxin and FlagOS complete MiniCPM5-2B deployment adaptation (source)openEuler · openEuler and Kunlunxin report P800 compatibility certification (source)Kunlunxin · Kunlunxin completes RLinf adaptation validation (source)Kunlunxin · AI accelerator portfolio (source)RLinf · RLinf Kunlunxin hardware support (source)Critical facts
IDC data reported by Reuters put Kunlunxin at 4% of China's data-center AI-accelerator shipments in the first half of 2026; unit volume, product mix, customers and workloads were not disclosed.
As of 2026-06-30Reuters · Nvidia's China AI-chip share falls as Huawei closes gap, report finds (source)Kunlunxin reports P800 validation across single- and multi-node RLinf modes, with a standard XPU plugin, XCCL collectives and Pi 0.5 LIBERO reproduction; release-v0.4 exposes setup instructions. Reported task and synchronization results lack a matched foreign-hardware control or independent reproduction.
As of 2026-10-10Kunlunxin · Kunlunxin completes RLinf adaptation validation (source)RLinf · RLinf Kunlunxin hardware support (source)Kunlunxin and FlagOS report Day-0 adaptation and inference validation, with a BF16 deployment image published through ModelScope and Hugging Face.
As of 2026-09-10Kunlunxin · Kunlunxin and FlagOS complete MiniCPM5-2B deployment adaptation (source)openEuler reports all seven compatibility subtests passed, covering device identification, driver and SDK deployment, and host/device and device/device links. This is partner-reported runtime compatibility, not independent AI workload performance, reliability at scale or CUDA parity.
As of 2026-10-19openEuler · openEuler and Kunlunxin report P800 compatibility certification (source)Company news
Timeline →openEuler and Kunlunxin report P800 compatibility certification
openEuler announced P800 compatibility certification on October 19; Kunlunxin reposted it October 23. The partners report seven passing tests of the basic operating environment.
Why it matters. This adds operating-system integration evidence. No matched training or inference result or production-scale reliability measurement is established.
openEuler · openEuler and Kunlunxin report P800 compatibility certification (source)Domestic accelerators gain reported share in China's first-half market
Reuters reports IDC data showing 2.008 million AI accelerators shipped in China in the first half of 2026. Huawei reportedly shipped 572,000 units for 28% share, behind Nvidia's 678,000 units and 34%. T-Head held 7%, Cambricon 5% and Kunlunxin 4%. Non-GPU accelerator shipments rose nearly 80% to 1.016 million units and exceeded GPU shipments by volume; Huawei, T-Head, Kunlunxin and Cambricon together accounted for 95% of that segment.
Why it matters. The figures are useful third-party market-adoption evidence and show domestic substitution at meaningful scale. They do not identify product mix, buyers, prices, accepted deliveries, utilization, workload performance, memory or fabrication provenance, and IDC's definition of an accelerator can include products aimed at different tasks. Nvidia's accessible product mix also differs under export controls, so market share is not a latest-versus-latest performance result.
Reuters · Nvidia's China AI-chip share falls as Huawei closes gap, report finds (source)Kunlunxin adds an RLinf backend for embodied reinforcement learning
Kunlunxin says its P800 completed RLinf validation across single-node, multi-node, training-simulation separation and training-inference separation. The implementation uses a standard XPU plugin, Ray scheduling and XCCL collectives, and the company reports Pi 0.5 training reproduction on LIBERO in eight- and 16-card configurations. RLinf's release-v0.4 repository exposes Kunlunxin hardware instructions alongside its other accelerator backends.
Why it matters. This is inspectable software-compatibility evidence and covers distributed execution beyond a single-card demo. The success-rate and synchronization figures are vendor tests without a matched foreign-hardware control, disclosed power or cost, independent reproduction, or evidence of production customer use. Compatibility with one reinforcement-learning stack does not establish broad framework coverage or CUDA parity.
Kunlunxin · Kunlunxin completes RLinf adaptation validation (source)RLinf · RLinf Kunlunxin hardware support (source)Kunlunxin and FlagOS publish a MiniCPM5-2B deployment image
Kunlunxin says it completed MiniCPM5-2B adaptation and inference validation with FlagOS and published a BF16 deployment image through ModelScope and Hugging Face.
Why it matters. The downloadable image is concrete compatibility evidence. It does not supply reproducible throughput, latency, power, accuracy or foreign-hardware comparison data.
Kunlunxin · Kunlunxin and FlagOS complete MiniCPM5-2B deployment adaptation (source)