Chinese AI labs

Chinese AI labs collectively became, between 2023 and 2026, the dominant force in open-weight frontier models. Per aggregator statistics reported in mid-2026, Chinese open models served roughly 61% of all tokens on OpenRouter, four of the five most-used models were Chinese, and Meta AI's Llama, the open-weight leader two years earlier, had fallen out of the top rankings entirely. This page maps every Chinese organization with large relevant pretraining capacity; each has a dedicated article.¹ Each organization's stated mission, quoted from its own materials, is tabulated alongside its Western counterparts at Stated missions of AI labs.

The landscape, as of July 2026

Lab	Type / backer	Flagship line	Openness	Distinguishing fact
DeepSeek	Hedge-fund-backed lab (High-Flyer)	DeepSeek V3/R1/V4	Open weights	Set the efficiency-disclosure norm; R1 shock repriced the industry
Qwen team	Alibaba (Tongyi lab)	Qwen 1 to Qwen 3	Mostly Apache 2.0	Most-downloaded base family on Hugging Face; passed Llama in cumulative downloads
Moonshot AI	Startup (Alibaba-backed)	Kimi K2/K3	Open weights	K3 (July 2026): ~2.8T parameters, the largest open-weight release to date, per reporting
Zhipu AI	Startup, Tsinghua spinout; HKEX-listed 2026	GLM 4.x/5.x	Permissive open	Longest-running open lineage; entity-listed 2025; first tiger to IPO
Tencent	Tech giant	Hunyuan (text/video/3D)	Increasingly open (Hy3 Apache 2.0)	WeChat-scale distribution; integrated DeepSeek-R1 rather than fighting it
ByteDance	Tech giant	Doubao / Seed lines	Mostly closed + Seed-OSS	China's largest assistant user base; world-class media models (Seedream/Seedance)
Baidu	Tech giant	ERNIE 4.x/5.x	4.5 open; 5.x API-only	First mover (ERNIE Bot, March 2023); in-house Kunlun chips
Meituan	Tech giant (local services)	LongCat	Open weights	LongCat-2.0 (1.6T) trained entirely on Chinese chips, per 2026 reporting
Ant Group	Alibaba-affiliated fintech	Ling / Ring (inclusionAI)	Open weights	Published frontier MoE training on mixed domestic silicon (2025)
Xiaomi	Consumer electronics	MiMo	Open weights	Reported ~21% of OpenRouter weekly tokens, Q2 2026; device-fleet strategy
MiniMax	Startup (Tencent/Alibaba-backed)	MiniMax-01 / M1 / M2	Open weights	Linear "lightning attention" at frontier scale; Hailuo video; Talkie apps
StepFun	Startup	Step 2/3 series	Mixed	Early trillion-parameter MoE; multimodal specialization
Huawei	Hardware champion	PanGu 5.x	Closed	Ascend chips: the domestic NVIDIA alternative; CloudMatrix racks
Shanghai AI Laboratory	State institute	InternLM / InternVL	Fully open	Academic anchor; OpenCompass evaluation platform; strongest open VLM line
iFlytek	Public company (speech)	Spark (Xinghuo)	Closed	Trained entirely on Huawei Ascend; entity-listed since 2019
SenseTime	Public company (vision)	SenseNova	Mixed	Shanghai AIDC compute-as-a-service; vision-era incumbent
Baichuan	Startup	Baichuan 2-4	Early open, now closed	Early open pioneer; pivoted to medical AI
01.AI	Startup (Kai-Fu Lee)	Yi series	Open (historical)	Exited frontier pretraining after R1; builds on DeepSeek

Smaller pretraining efforts with real but sub-frontier capacity: ModelBest/OpenBMB (Tsinghua-affiliated; the MiniCPM efficient-model line), Kunlun Tech (Skywork models and widely used open reward models), and Xiaohongshu/RedNote (the dots.llm1 open MoE). These are tracked in the registry and covered at mention level pending dedicated articles.¹

How the landscape formed

Three phases structure the story. The hundred-model war (2023): after ChatGPT, incumbents (Baidu, Tencent, ByteDance) and a venture cohort, the "AI tigers" (Zhipu AI, Moonshot AI, MiniMax, StepFun, Baichuan, 01.AI), launched rival assistants under China's interim generative-AI measures, with LLaMA-descended and homegrown bases proliferating. The efficiency turn (2024-2025): DeepSeek's open reports, then DeepSeek-R1, reset the game around disclosed, cheap, open frontier training; capital consolidated, Baichuan and 01.AI pivoted away from the frontier, and open weights became the sector's export product. The domestic-stack ascendancy (2025-2026): Qwen took the ecosystem crown from Llama, usage aggregators tipped toward Chinese models, trillion-parameter open releases arrived at sub-quarter cadence (Kimi K2/K3, LongCat, DeepSeek V4), and training migrated visibly toward Chinese silicon at Huawei, Ant Group, Meituan, and iFlytek.¹²

Cross-cutting facts

Compute and export controls. Every lab operates under the US export-control regime documented on this wiki's governance pages; responses span stockpiled and compliant NVIDIA parts (ByteDance, the tigers), in-house silicon (Baidu's Kunlun, Huawei's Ascend), and mixed-fleet engineering (Ant Group). The 2026 all-domestic training runs are the regime's sharpest counter-evidence to date.²

Disclosure culture. Chinese leaders publish more than their Western counterparts: full technical reports with data pipelines, RL recipes (RLVR-style), and cost accounting are the norm from DeepSeek, Moonshot, Zhipu, and the Qwen team, inverting the closed turn taken by OpenAI and peers after GPT-4.¹

Open-weight economics. Flagship-class open releases priced near cost, subsidized by cloud, device, or platform businesses, are the sector's competitive weapon, forcing the pricing responses documented on Western model pages from Claude Opus 4.5 to the GPT-5.x tiers.¹³

References