Xiyu Technology on Feb. 13 open‑sourced its new flagship large language model, MiniMax M2.5, and within hours the model had been adapted and deployed on Huawei’s Ascend‑based Atlas 800 A2 and A3 servers. Huawei’s computing account said the rapid porting and immediate trials at multiple live network nodes demonstrate full‑process compute support for scaling the model across operational environments.
The speed of the deployment reflects close coupling between model development and hardware optimisation. Ascend — Huawei’s in‑house AI chip family — together with the company’s AI Agent orchestration capabilities, was credited with enabling the model to run on Atlas 800 racks quickly, a signal that vendors and model teams are prioritising co‑design to shorten the path from research to production.
Open‑sourcing MiniMax M2.5 matters beyond a single model release. A permissive code base lowers the barrier for domestic developers and service providers to experiment, optimise and integrate the model into applications, while support for Ascend accelerators offers a route to reduce dependence on foreign GPUs and their supply chains.
Huawei’s mention of trials at “现网局点” — operational network sites — is notable because it suggests real‑world, operator‑grade testing rather than lab demonstrations. That carries implications for latency‑sensitive applications such as edge inference, telecoms automation and large‑scale conversational services that require integration with existing infrastructure.
The announcement also fits a broader trend in China: public and private actors are building a vertically integrated AI stack, from models to chips to cloud and telco deployments. For domestic cloud providers, chip designers and model creators, the ability to certify and operate LLMs on local hardware is becoming a strategic priority amid export controls and geopolitical frictions that complicate reliance on Western compute ecosystems.
Caveats remain. The statement did not include independent benchmarks, cost metrics or details about power consumption and throughput, so performance claims are yet to be verified by third parties. Open‑sourcing increases adoption but also raises questions about safety, misuse and regulatory oversight as more capable models leave research silos.
Looking ahead, expect further model releases and tighter optimisation between Chinese LLMs and Ascend hardware, plus deeper partnerships between vendors and telecom operators. Observers should watch for published benchmarks, broader rollouts beyond trial nodes, and how regulators and enterprise customers respond to the accelerating domestic AI stack.