Can China’s Homegrown LLMs Compete After DeepSeek’s Open‑Source Disruption?

In early 2025 DeepSeek R1 was open‑sourced under the MIT License, an event the article describes as a bomb that triggered a chain reaction across the global large‑model landscape.

Immediate industry reactions

OpenAI hurriedly released o3‑mini, and its CEO admitted the company was "on the wrong side of history".

French leader Mistral AI launched the new generation application Le Chat, calling DeepSeek a catalyst that lowers cost and breaks the traditional compute‑arms race.

Analyst reports from Deutsche Bank and various broker research projects predict DeepSeek’s open‑source strategy will reshuffle the "hundred‑model battle" toward a few dominant players.

Domestic cloud and platform response

Major Chinese cloud providers—including Huawei Cloud, Alibaba Cloud, Tencent Cloud, Baidu Cloud, JD Cloud and Unicom Cloud—rolled out joint services, zero‑code deployment options, and limited‑time free trials. Platforms such as OpenRouter and MiTa AI Search integrated DeepSeek, while enterprises like Visual China, Xiaomi and news‑aggregation services began using the model.

Technical advantages of DeepSeek R1

The MIT License permits free commercial use, modification without royalties, and the ability to keep derived code closed, encouraging developers to build plugins and extensions. DeepSeek R1 combines strong inference and reflection abilities with knowledge‑distillation techniques that produce 32B, 14B, 7B and 1.5B variants capable of running on edge devices, reducing data‑transfer costs while preserving security. Its multi‑scenario adaptability lets it run from cloud to local CPUs at roughly 3% of the cost of comparable OpenAI models, yet achieving performance comparable to GPT‑4o.

Ecosystem growth

The open‑source model quickly fostered a large ecosystem: global developers contribute code, bugs are fixed rapidly, hardware vendors accelerate chips to meet compute demand, and new business models—hosted services, custom plugins, vertical fine‑tuning—emerge like mushrooms.

The "only first, no second" reality

The article argues that early market share grants massive resources, creating a short‑distance sprint where the first model to dominate captures most honors and funding. Models that cannot match DeepSeek’s cost‑performance fall behind, leading to a concentration of resources on the leading model.

Importance of ecosystem construction

Open‑source drives network effects; newcomers or lower‑ranked models face huge barriers to break the ecosystem wall.

Emerging technical trends

Chain‑of‑thought, parameter distillation, mixture‑of‑experts (MoE) and multi‑head latent attention are identified as key challenges. Cloud platforms now provide one‑click deployment and acceleration for DeepSeek series, lowering entry thresholds for applications in finance, smart manufacturing, and more. Cross‑domain data sharing, shared training resources, and feedback‑driven continuous iteration further support complex, variable scenarios.

How Chinese models might break through

Alibaba Tongyi Qianwen : Leverages Alibaba Cloud’s market dominance and DAMO Academy’s AI research to offer customized enterprise solutions and improve MoE‑Transformer architectures.

ByteDance Doubao : Integrates with Douyin and other products to provide multimodal generation, such as video creation assistants for creators.

Kimi : Focuses on reinforcement learning to refine reward models and targets the education sector with intelligent tutoring systems.

MinMax : Pursues multimodal research, aims for infinite input/output length, and expands to overseas markets with a platform already active in 20 countries.

Zhipu AI : Continues scaling its Step series from billions to trillions of parameters, covering language and multimodal capabilities.

iFlytek Spark : Leverages its speech‑AI strengths and extensive industry partnerships to fuse voice and large‑model technologies for education, healthcare, and other domains.

Face‑wall MiniCPM : Targets edge AI with long‑text and high‑resolution image processing, adding real‑time video understanding for smart devices.

The article concludes by inviting community suggestions on how domestic models can break through the DeepSeek‑driven disruption.