On June 12, 2026, Richard Yu's HDC 2026 keynote in Dongguan officially unveiled openPangu 2.0. Eighteen days later, on June 30, openPangu-2.0-Flash model weights, base inference code, and training/inference operators went live on GitCode. This is not just another weight drop—it is the first frontier-scale open-source LLM trained entirely without NVIDIA hardware, and one of the few ultra-large MoE models offering genuine full-stack open source.

Two variants, one 512K context window:

512K context means roughly eight full-length novels in a single pass—entire contracts, large codebases, and multi-hour conversation histories fit without chunking.

Seven components on the open-source roadmap:

openPangu 2.0 uses a Mixture-of-Experts (MoE) architecture. Key technical components:

Trained without a single NVIDIA GPU: All training ran on Huawei Ascend 910B NPUs—zero A100 or H100 usage. With US export controls tightening on advanced AI chips to China, Huawei not only trained a 505B MoE model but also achieved:

Developer stack: Built on CANN (Huawei's CUDA-class runtime) plus torch_npu (PyTorch adapter). Standard PyTorch code switches to Ascend with import torch_npu. Deployment paths: Huawei Cloud ModelArts API, GitCode Ascend Tribe self-host, and HarmonyOS native on-device integration.

A Flash-Int8 quantized build is already available with W4A8 quantization: 40% less memory, accuracy loss under 10%.

import torch
import torch_npu

model = load_openpangu("./openPangu-Flash")
model = model.to("npu:0")
output = model.generate(input_ids.to("npu:0"), max_new_tokens=512, temperature=0.7)

Capability matrix (architecture-based inference; independent third-party benchmarks pending):

openPangu 2.0 is not the overall strongest open model today—DeepSeek V4 Pro still leads on code and hard reasoning. But on 512K context, sovereign AI supply chains, native Ascend 2× throughput, full-stack open source, and on-device deployment, it occupies a category of one.

Geopolitical significance: openPangu 2.0 is the first frontier-scale open-source model trained without NVIDIA hardware. At HDC 2026, Richard Yu stated: "In my dictionary for the rest of my life, there is no second place—only first." Full-stack open source matters for academic reproducibility, vertical-domain retraining, and Ascend ecosystem growth under export-control pressure.

HarmonyOS Agent foundation: HarmonyOS 7 enters the Agent era with openPangu 2.0 as the native AI engine. HarmonyOS Agent Framework 2.0 reports >90% success on complex multi-step tasks. The on-device 30B model runs locally on phones without network access.

openPangu License: Commercial use permitted, royalty-free, non-exclusive. Confirm exact terms in the GitCode repository.

Open-source roadmap:

API-only access cannot replace a production agent host: laptop lids close, multi-key management gets messy, and HarmonyOS/OpenClaw pipelines need macOS always online—local debugging and cloud API each carry hidden costs. For teams running 7×24 multi-model agents that route across Huawei Cloud ModelArts and OpenRouter, a KVMNODE dedicated cloud Mac Mini is usually the steadier path: native Apple Silicon toolchains, daily/weekly/monthly terms. See the pricing page; order via the order page.