Topics

China AI company watchlist (which companies deserve weekly checks)

Evergreen topic pages updated with new evidence

Last reviewed: 2026-05-13 · Policy: Editorial standards · Methodology

Decision in 20 seconds

A useful China AI company watchlist in 2026 focuses on three tiers: foundation model labs (DeepSeek, Qwen/Alibaba, Moonshot, Zhipu, MiniMax, StepFun) whose releases directly affect what you can build; workflow infrastructure companies (Dify, Coze, LobeChat) that determine how Chinese AI is packaged for production; and physical-AI players (Momenta, WeRide, Unitree) whose deployment milestones are the leading indicators for real-world AI adoption. Ranking all companies on a single list optimizes for noise. The right structure separates these tiers by monitoring frequency: labs weekly, infra monthly, physical AI quarterly.

Use this page when

  • You need to decide which Chinese AI companies deserve recurring attention vs. one-time monitoring.
  • You want to separate foundation model lab signals from workflow infra and physical AI signals.
  • You're building a monitoring stack and need to allocate attention budget across tiers.

This page is not for

  • Comprehensive China AI company database — this is a curated builder-first watchlist, not a directory.
  • Investment or financial analysis — funding figures are indicative, not investment advice.
  • Consumer-facing AI product rankings — this page focuses on infrastructure and B2B layers.

Key points

  • Foundation model labs: DeepSeek (MIT license, cost-efficient API), Qwen/Alibaba (Apache 2.0, Qwen3 April 2026), Moonshot/Kimi (long-context agentic), Zhipu (GLM, enterprise function-calling), MiniMax (1M context), StepFun (multimodal). These 6 define the capability surface.
  • Workflow infra: Dify (open-source LLM app framework, 50k+ GitHub stars by 2025, Series B 2024), Coze (ByteDance, commercial AI app builder, international expansion 2025), LobeChat (open-source, self-hosted AI chat platform).
  • Physical AI: Momenta (L4 AV, deployed in Wuhan and Shenzhen robotaxi fleets 2024–2025, Series C $500M+), WeRide (NYSE: WRD, commercial driverless operations in 30+ cities by 2025), Unitree (humanoid robot G1 series, 2025 production ramp).
  • Funding signal: $500M+ rounds in 2024–2025 went to Moonshot ($1B Series B), MiniMax (Alibaba-led round), and DeepSeek (Kuaishou-backed). Large rounds don't automatically mean capability lead — check benchmarks.
  • China's model labs are releasing on Apache 2.0 / MIT licenses at increasing frequency — this shifts the on-prem deployment math for builders who previously had to pay enterprise licensing.
  • Dify's success (commercial traction, Series B, international adoption) shows the layer between model APIs and end-user applications is where packaging value accrues in the Chinese AI stack.

What changed recently

  • April 2026: Qwen3 Apache 2.0 release — foundation model tier now has two fully permissive options (Qwen3 + DeepSeek MIT).
  • March 2026: Momenta expanded robotaxi operations to third city (Hangzhou); now operating Wuhan, Shenzhen, Hangzhou with commercial fares.
  • February 2026: Dify v1.0 release with visual workflow builder; enterprise tier adoption accelerated in US and EU markets.
  • January 2026: StepFun raised Series B at $3B valuation, positioning as multimodal specialist distinct from pure-text labs.
  • Late 2025: Unitree G1 humanoid robot entered small-batch production; pricing at ~$16,000 for commercial buyers.

Explanation

The most common mistake with a company watchlist is trying to rank companies 1–10. This collapses three very different monitoring cadences. Foundation model labs change your build options weekly; workflow infrastructure changes your deployment options monthly; physical AI moves quarterly at best.

Chinese AI company monitoring requires separating company-level signals from model-level signals. A model release from Alibaba (Qwen3) doesn't mean you should watch Alibaba Cloud enterprise packaging as frequently — these are different monitoring questions.

For builders, the most actionable watchlist signals are: (1) new Apache 2.0 or MIT model from a tier-1 lab → evaluate for on-prem deployment; (2) Dify/Coze major release → re-evaluate app framework choice; (3) Momenta/WeRide new city deployment → early indicator of physical AI infrastructure readiness.

China AI company watchlist — by tier and monitoring frequency

Monitoring frequency differs by tier. Collapsing all companies into one weekly read wastes attention on slow-moving signals.

How to verify the answer

For model releases, check official GitHub/HuggingFace pages. For company-level news (funding, product), 36Kr Global and KR Asia are the most reliable English wires.

Tools / Examples

  • Dify Series B (2024): Dify's commercial round validated that the open-source LLM app framework layer had product-market fit beyond developer experimentation. Builders using LangChain re-evaluated.
  • DeepSeek V3 MIT license (Nov 2024): Changed the on-prem deployment calculus for Chinese enterprises and international builders simultaneously — first Chinese tier-1 model with no usage cap under MIT.
  • Momenta Wuhan robotaxi commercialization (2024): First Chinese AV company to charge commercial fares in a major city — a physical-AI deployment milestone that preceded the headline funding rounds by 6 months.

Evidence timeline

Sources

FAQ

What are the most important China AI companies to watch in 2026?

By tier: Foundation models — DeepSeek, Qwen (Alibaba), Moonshot (Kimi), Zhipu (GLM), MiniMax, StepFun. Workflow infra — Dify, Coze (ByteDance). Physical AI — Momenta, WeRide, Unitree. These 11 companies cover the most actionable signals for builders.

How is the China AI company watchlist different from the model release tracker?

The model tracker follows individual model releases (benchmarks, licenses, API pricing). The company watchlist tracks which organizations are worth recurring attention — their funding, team signals, product direction, and deployment milestones that don't always show up in model release notes.

Which Chinese AI companies have raised the most capital?

Moonshot/Kimi ($1B Series B, 2024), MiniMax (Alibaba-led round, ~$600M), DeepSeek (Kuaishou-backed, exact terms not public), Momenta ($500M+ Series C). StepFun ($3B valuation at Series B, 2026). Note: fundraising size correlates weakly with model capability — check benchmarks independently.

Is Dify worth following for builders outside China?

Yes — Dify is the most internationally adopted Chinese AI infrastructure company. It's open-source (GitHub stars: 50k+ by 2025), has a commercial cloud tier, and is used by teams in the US and EU for LLM app orchestration. Series B confirmed commercial traction.

When should a company leave the watchlist?

Remove a company when: (1) no significant update (model, product, funding) in 12+ months; (2) key team departures with no successor signals; (3) model performance benchmarks consistently below watchlist peers with no differentiated use case.

Should I watch Baidu and Alibaba on this list?

Baidu (ERNIE) and Alibaba (Qwen) are better tracked through their model families than at the company level — the relevant signals are model releases, API pricing changes, and license terms. Alibaba as a company is too broad for a weekly watchlist; Qwen as a model family is the right unit.

What monitoring frequency do these companies warrant?

Weekly: foundation model labs (DeepSeek, Qwen, Moonshot, Zhipu, MiniMax, StepFun) — model releases can be actionable immediately. Monthly: workflow infra (Dify, Coze) — product releases affect deployment choices but less urgently. Quarterly: physical AI (Momenta, WeRide, Unitree) — city deployments and hardware milestones move slowly.

How do I track these companies without spending 2 hours a day?

Use RadarAI weekly digest for routing (15 min/week). Watch QwenLM GitHub and DeepSeek HuggingFace for model releases (5 min/week). Subscribe to Dify GitHub releases for infrastructure updates. Check 36Kr Global for funding/distribution news on your monthly review.

Search angles this page supports

Related

Go deeper

Last updated: 2026-05-13 · Policy: Editorial standards · Methodology