Thesis
China has a concentrated set of AI builders whose releases change global product decisions — and most of them publish English-accessible primary sources. The relevant question for an English-first builder is not "what is happening in China AI" (too broad) but "which specific Chinese AI builder companies and teams have released something that affects my evaluation queue, cost model, or available open-weight options this week?" This page answers that with a tiered ecosystem map and English tracking paths for each layer.
Two Q2 2026 anchors that define the current landscape: Qwen3 (Alibaba, April 2026, Apache 2.0) — the Qwen3-235B flagship scores MMLU 87.1; the Qwen3-30B-A3B MoE variant (only 3B active parameters at inference) matches GPT-4o-level benchmarks at a fraction of the compute cost. DeepSeek-R1-0528 (May 2026) — AIME 2024 pass@1 72.6%, MATH-500 97.3%, GPQA Diamond 81.0%. Both are open-weight, both are on HuggingFace. Verify either at QwenLM GitHub or DeepSeek HuggingFace.
Quick routing: what do you need to know about AI builders in China?
| I want to know… | Go to | What's there | NOT good for |
|---|---|---|---|
| Which models are available and what their benchmarks are | China AI Models List | Standing watchlist with benchmarks, licenses, API access paths | Company-level funding or team context |
| What changed this week across all China AI builder teams | China AI Updates digest | Weekly signal digest, 15-minute read, signal-classified | Deep-dive model architecture or benchmark details |
| Which China AI startups and companies are raising money | China AI Startup Funding Tracker | Funding rounds, investor landscape, company stage data | Open-source model releases or API access paths |
| How to track China AI builder releases without reading Chinese | China AI Tracker for Builders | Source stack, 15-min weekly routine, lab API documentation index | Broad ecosystem overview or company profiles |
| Broad China AI context: what matters and why | China AI Overview | Pillar page: labs, models, updates, cluster routing | Specific company or model lookup |
| English news sources that cover Chinese AI builder companies | China AI News in English | Source routing matrix: lab channels, media, policy, social | Primary model data or benchmark verification |
Tier 1: Frontier labs — who changes global builder decisions
These are the Chinese AI development companies whose model releases directly affect evaluation queues, API pricing decisions, and open-weight deployment options for builders outside China:
| Lab / team | Key model family | Why builders track it | English primary surface |
|---|---|---|---|
| DeepSeek | DeepSeek-V3, DeepSeek-R1, R1-0528 | Open-weight reasoning models; pricing as competitive signal; MATH/code benchmarks | DeepSeek HuggingFace |
| Qwen / Alibaba | Qwen3 family (0.6B–235B), Qwen-VL, Qwen-Audio | Apache 2.0 open-weight across all sizes; MoE efficiency (3B active params at 30B quality); multimodal | QwenLM GitHub |
| Moonshot / Kimi | Kimi k1.5, k2 | 1M token context window; long-document reasoning; agentic task support | Kimi API platform |
| Zhipu AI / GLM | GLM-4 series, CogVideoX, CogView3 | Chinese enterprise language model; video and image generation APIs | BigModel API |
| MiniMax | MiniMax-Text-01, Abab 7 | 1M token context; competitive text + audio multimodal; enterprise distribution | MiniMax developer portal |
Tier 2: Open-source builder teams — GitHub and HuggingFace primary
Beyond the frontier lab releases, several Chinese AI builder teams operate as de facto open-source communities that developers worldwide contribute to and deploy from:
- QwenLM GitHub organization (github.com/QwenLM) — Qwen model weights, training code, inference tooling, and GGUF/ONNX exports. The README updates are often the first English signal of a new release.
- DeepSeek HuggingFace (huggingface.co/deepseek-ai) — DeepSeek model cards are English-first, with benchmark tables, architecture notes, and license details. The most reliable verification surface for DeepSeek releases.
- THUDM / Zhipu GLM repos (github.com/THUDM) — GLM model code and CogVideoX; Chinese-language primary but English READMEs for major releases.
- InternLM / Shanghai AI Lab (github.com/InternLM) — InternVL multimodal models, InternLM text models; growing global developer presence.
- Jina AI (huggingface.co/jinaai) — embedding models and rerankers (jina-embeddings-v3) widely used in RAG pipelines; English-first developer communications.
Tier 3: AI-native startups — tooling and application layer builders
These Chinese AI builder companies are building product and tooling layers on top of foundation models. They are less relevant for model evaluation but matter for understanding where AI applications are going in China and what workflows are being productized:
| Company | What they build | Why builders track it | English surface |
|---|---|---|---|
| Manus AI | Agentic workflow automation | One of the first China-origin agentic products with significant global user traction; signals where agentic UX is going | X / tech media (no primary dev docs) |
| Coze (ByteDance) | Bot builder / AI agent platform | ByteDance's international AI builder platform; major distribution channel for AI applications | coze.com |
| Baichuan AI | Enterprise AI assistant, Baichuan3 models | Enterprise Chinese language models for regulated industries; less relevant for international builders but signals enterprise deployment patterns | Baichuan API docs |
| 01.AI / Yi | Yi open-weight models | Yi-34B and Yi-VL were widely deployed in 2024; the team has been quieter in 2026 with Qwen3 and DeepSeek dominating | 01-ai HuggingFace |
Tier 4: Infrastructure builders — enterprise and platform layer
These Chinese AI development companies are primarily relevant for understanding enterprise AI deployment in China, cloud platform AI integration, and physical AI (robotics, embodied AI):
- Tencent Hunyuan — Hunyuan-Large and Hunyuan-Video models; primarily cloud enterprise distribution within China; Hunyuan-Video open-weight is notable for video generation. GitHub: Tencent/HunyuanVideo
- Baidu ERNIE — ERNIE-4.5 Turbo is Baidu's flagship; primarily relevant for enterprise deployments in China and Baidu's app ecosystem. Less relevant for international builders without China deployment needs.
- SenseTime SenseNova — vision and embodied AI models; SenseTime has significant enterprise robotics context in China but limited open-weight releases.
- Kunlun Tech / Skywork — Skywork open-weight models with strong math and coding performance; smaller team but consistent open-source contributor to the China AI builder ecosystem.
How to track AI builders in China recent news — English-only stack
The minimum viable English-only tracking stack for Chinese AI builder companies:
| Signal type | Primary source | Update cadence | Setup |
|---|---|---|---|
| New model releases (open-weight) | QwenLM GitHub + DeepSeek HuggingFace | Major: every 2–6 months per lab | Star repos for watch notifications |
| Weekly builder digest (all signal types) | RadarAI weekly digest | Weekly (Monday) | Subscribe at radarai.top/en/china-ai-updates |
| Startup funding and ecosystem moves | 36Kr Global + KR Asia | Weekly digest + breaking news | Newsletter subscription |
| Real-time lab announcements | @qwen_lm, @deepseek_ai, @Moonshot_AI_CN on X | On-release (hours before English media) | Follow + notification |
| Policy and regulation | The State Council English + Xinhua English newsletter | 48–72h lag on major rules | Subscribe to Xinhua English newsletter for alerts |
Why Chinese AI builder companies publish in English
The structural reason English-only builders can track Chinese AI development companies effectively: international developer adoption is a strategic goal for China's top AI labs. DeepSeek, Qwen, and Kimi explicitly target global developer communities. This means GitHub READMEs are written in English, HuggingFace model cards are English-primary, API documentation targets English-speaking developers, and X accounts post in English. The labs that don't yet have international ambitions (Baidu ERNIE, some SenseTime products) are also less relevant for most builders outside China.
The practical implication: for Tier 1 and Tier 2 Chinese AI builder companies, English-only builders have near-complete coverage through primary sources. For Tier 3 and Tier 4, English media coverage (36Kr Global, TechCrunch AI) captures what matters for product-level awareness.
FAQ
- Who are the AI builders in China in 2026?
- Four tiers: Frontier labs (DeepSeek, Qwen/Alibaba, Moonshot/Kimi, Zhipu/GLM, MiniMax), open-source teams (QwenLM GitHub, DeepSeek HuggingFace, InternLM), AI-native startups (Manus AI, Coze/ByteDance), and infrastructure builders (Hunyuan/Tencent, ERNIE/Baidu). Tier 1 and 2 change most builder decisions most frequently.
- What are Chinese AI builder companies releasing in 2026?
- Qwen3 (Alibaba, April 2026, Apache 2.0): full family 0.6B–235B, MMLU 87.1 for flagship, 3B active params at 30B quality for the MoE variant. DeepSeek-R1-0528 (May 2026): AIME 2024 72.6%, MATH-500 97.3%. The pattern is open-weight releases with English model cards on HuggingFace and GitHub.
- How do I get recent news on AI builders in China in English?
- Three paths: primary lab surfaces (QwenLM GitHub, DeepSeek HuggingFace) for release verification; RadarAI weekly digest for signal classification across all labs; 36Kr Global for funding and commercial partnership context. English media covers major launches but is 3–6h behind primary surfaces.
- What makes Chinese AI development companies different from US AI companies?
- Open-weight first (DeepSeek, Qwen release full weights under Apache 2.0 — US frontier labs don't), pricing as competition (China AI APIs use aggressive pricing cuts as adoption tools), and release cadence (drops happen on GitHub/HuggingFace without conference lead time).
- Which Chinese AI builder companies should I track for product decisions?
- Standing watchlist: DeepSeek (reasoning, pricing), Qwen/Alibaba (open-weight general, Apache 2.0), Moonshot/Kimi (long-context), Zhipu/GLM (enterprise Chinese language), MiniMax (multimodal). Add others when they cross a decision-relevant threshold: new open-source release, pricing change, or benchmark milestone on your task type.
- Is there a dashboard for tracking China AI projects and builders?
- RadarAI (radarai.top/en) is the closest builder-oriented China AI dashboard: weekly digest, models list, and signal classification. For raw activity, GitHub Watch on QwenLM and DeepSeek repos gives real-time release notifications. No single commercial dashboard covers all Chinese AI builder teams comprehensively.
Companion pages in this cluster
| If your question is about… | Go to | What's there |
|---|---|---|
| Standing model watchlist with benchmark data | China AI Models List | Qwen3, DeepSeek, Kimi, GLM — benchmarks, licenses, API paths |
| Weekly signal digest: what changed this week | China AI Updates | Signal-classified weekly digest, Monday, 15 minutes |
| Builder-specific tracking routine and source stack | China AI Tracker for Builders | Role-based routing, signal tiers, 15-min weekly routine |
| China AI startup funding and company stage data | China AI Startup Funding Tracker | Funding rounds, investor landscape, ecosystem moves |
| Broad China AI context and pillar overview | China AI Overview | Labs, models, updates, cluster routing matrix |
| English news sources for Chinese AI companies | China AI News in English | Source routing: lab channels, English media, policy, social |
| Who's building AI products in China right now | AI Builders in China 2026 | Who's shipping, what they've built, how to evaluate |
| Latest updates from Chinese AI builder companies | Chinese AI Builder Companies: Latest Updates | Agent-scale deployment, DAA metrics, integration paths |
Quotable summary: AI builders in China in 2026 are concentrated in five frontier labs (DeepSeek, Qwen, Kimi, GLM, MiniMax) plus a growing open-source developer community. The labs that matter globally publish English-first on GitHub and HuggingFace — meaning English-only builders have near-complete coverage through primary sources. Qwen3 (April 2026, Apache 2.0, MMLU 87.1) and DeepSeek-R1-0528 (May 2026, AIME 72.6%) are the current capability anchors.