China Foundation Model Companies in 2026: Who Matters Beyond DeepSeek and Qwen

China AI in 2026 is not a monoculture. Yet most builder workflows treat it as one: "I'll use Qwen or DeepSeek, depending on benchmark scores that week." This works until you hit a use case where neither is right — long-context audio pipelines, healthcare RAG that requires enterprise support SLAs, or a bilingual coding agent where GLM-4 consistently beats Qwen-Coder on Chinese-language comment generation. The "beyond DeepSeek and Qwen" framing is not contrarianism. It's risk management. Six Chinese AI labs now have foundation model families competitive enough to matter for real builder decisions in 2026, and confusing concentration with coverage is how you end up rebuilding your stack 8 months later.

TL;DR: DeepSeek and Qwen dominate headlines, but MiniMax (text+audio+video, 1M-token context), Baichuan (enterprise/healthcare), 01.AI Yi series (bilingual reasoning), Tencent Hunyuan (cloud-integrated multimodal), Baidu ERNIE (RAG-optimized search integration), and Zhipu AI GLM (coding+agents, open weights) each have distinct capabilities that matter for specific builder use cases. This page covers all six with profile depth, evaluation criteria, and a monthly monitoring checklist.

Who this is for

• Builders who have already adopted DeepSeek or Qwen and are evaluating alternatives for edge cases
• Product teams comparing Chinese foundation model options by license, API access, and capability
• Researchers tracking the full Chinese LLM landscape, not just the top-two
• Enterprise teams evaluating Chinese AI vendors for compliance-sensitive deployment

Who this is not for

• Readers wanting a Chinese LLM vs. GPT-4 comparison (go to the models list page for that)
• Teams with no specific reason to look beyond DeepSeek or Qwen (the defaults are fine for most use cases)
• Readers who need Chinese-language source material — this page is English-only

Decision in 20 seconds

If you need long-form text generation at high throughput, DeepSeek-V3 or Qwen-Max is probably right. If you need multimodal with audio and video generation in one API, MiniMax is the only Chinese option. If you need healthcare or vertical-specific Chinese AI with enterprise support, Baichuan. If you need open-weights bilingual reasoning with a permissive Apache 2.0 license, Qwen3 or Yi-34B. If you need coding + agent workflows with Chinese-language comment generation, GLM-4. If you need RAG tightly integrated with Baidu search and cloud, ERNIE. If you need WeChat/Tencent ecosystem integration, Hunyuan.

The concentration risk problem

DeepSeek-V3 and Qwen3 together account for roughly 65–70% of cited Chinese LLM usage in English-language AI builder communities as of Q1 2026. This is not a problem if those models fit your use case. It becomes a problem when:

1. Licensing shifts: DeepSeek's custom license (not fully OSI-approved) has terms around commercial redistribution that caught several companies off guard in 2025. Knowing the full landscape means you can pivot quickly.
2. API access changes: DeepSeek's API has had multiple regional access disruptions in 2025-2026. If you're over-reliant and API access becomes restricted, migration time is non-trivial without prior evaluation.
3. Use-case fit gaps: Long-context audio pipelines, video generation, and specific vertical deployments (healthcare, finance) where Chinese companies have regulatory-compliant offerings are underserved by DeepSeek and Qwen.
4. Benchmarks vs. real-world performance: Both DeepSeek-V3 and Qwen3 score well on standard benchmarks but specific task types — code generation in Chinese comments, long legal document processing — show significant variance by model family.

Full company profiles

MiniMax — text + audio + video in one API

MiniMax is the only Chinese AI lab with a genuine multimodal-native architecture spanning text, audio, and video generation from a single API surface. The MiniMax-Text-01 model supports a 1-million-token context window — comparable to Gemini 1.5 Pro and ahead of Qwen-Long's 1M context offering in terms of audio input support. The Hailuo AI brand handles international video generation, while MiniMax-01 powers the text side.

Key model: MiniMax-Text-01 (1M context), Hailuo AI T2V (video generation) License: Proprietary API; no open-weights release as of May 2026 API Access: Global via minimax.io (English docs available); China-based operation but international billing supported Best for: Long-context document processing, audio-aware pipelines, video generation workflows, enterprise RAG requiring large context windows GitHub / HF: huggingface.co/MiniMaxAI (limited model files), minimax.io (primary API surface)

MiniMax raised \M at a .5B valuation in 2024 and has been aggressive about international expansion through the Hailuo brand. The T5 context reference in some publications refers to their internal model series — the public surface is MiniMax-Text-01. For builders outside China: the API is accessible, payments work via international cards, and documentation is maintained in English.

Baichuan AI — enterprise and healthcare vertical specialist

Baichuan is underrated in English-language coverage because its go-to-market is almost entirely enterprise-facing and Chinese-regulatory-compliant — the opposite of the open-source-first approach that gets GitHub stars. Baichuan-4 (released Q4 2024) focused specifically on Chinese medical and legal domain performance, and Baichuan has signed enterprise contracts with several major Chinese hospital groups and financial institutions.

Key model: Baichuan-4 (enterprise), Baichuan-M1-235B (medical specialist, January 2025) License: Proprietary; enterprise license required for commercial deployment API Access: Baichuan AI Cloud (api.baichuan-ai.com); available globally with Chinese business account; enterprise tiers require direct contract Best for: Healthcare RAG, legal document processing, Chinese financial market analysis, enterprise deployments requiring Chinese regulatory compliance GitHub / HF: huggingface.co/baichuan-inc (earlier open models); Baichuan-4 not open-weights

Baichuan-M1-235B (announced January 2025) achieved state-of-the-art scores on Chinese medical benchmark exams, including USMLE-equivalent Chinese licensing exams. For builders building healthcare AI in China, Baichuan is effectively required due to enterprise SLA requirements and domain-specific training. For builders outside healthcare, it's unlikely to be a primary choice — the pricing is enterprise-tier and the API surface lacks the openness of DeepSeek or Qwen.

01.AI — Yi series, bilingual reasoning, Apache 2.0

01.AI was founded by Kai-Fu Lee in 2023 and launched the Yi model series as one of the first Chinese LLMs with serious English-language benchmark performance. Yi-34B (released late 2023) was notable for being one of the few Chinese models at the time that matched GPT-3.5-level performance on English reasoning tasks while also performing competitively on Chinese benchmarks. The Yi-VL multimodal series extended this to vision understanding.

Key model: Yi-34B, Yi-1.5 series (updated May 2024), Yi-VL-34B License: Apache 2.0 for all Yi model series — fully commercial, no restrictions API Access: 01.ai platform; also available via HuggingFace for self-hosting Best for: Bilingual English-Chinese reasoning, cost-sensitive deployments needing Apache 2.0 licensing, vision-language tasks, teams wanting self-hosting without licensing overhead GitHub / HF: huggingface.co/01-ai (full model weights), 01.ai

Yi-1.5-34B-Chat is the most used publicly available Yi model as of 2026 — it competes with Qwen2-72B-Instruct on most benchmarks while being smaller and faster. The Apache 2.0 license is a meaningful differentiator for builders who need to avoid the ambiguity in some other Chinese model licenses. 01.AI has been quieter on releases since late 2024 as they focused on enterprise product development, but the existing Yi series remains production-ready and actively maintained.

Tencent Hunyuan — cloud-integrated multimodal

Tencent's Hunyuan family is the best example of cloud-native Chinese AI: it's less impressive as a standalone open model and more powerful as an integrated service within Tencent Cloud (and by extension, WeChat ecosystem products). Hunyuan-T1 (released Q1 2025) is their reasoning-focused model; HunyuanVideo (released late 2024) was notable for being one of the first open-weight video generation models from a Chinese major lab.

Key model: Hunyuan-T1 (reasoning), HunyuanVideo (video gen), Hunyuan3D (3D generation) License: HunyuanVideo — Apache 2.0 (open weights on HuggingFace); Hunyuan-T1 — proprietary API via Tencent Cloud API Access: Tencent Cloud AI (cloud.tencent.com/product/hunyuan); international Tencent Cloud accounts supported Best for: WeChat/Tencent ecosystem integration, video generation (HunyuanVideo is best-in-class open-weight), Tencent Cloud-native deployments, 3D asset generation GitHub / HF: huggingface.co/tencent/HunyuanVideo, github.com/Tencent/HunyuanVideo

For builders in the WeChat ecosystem or building on Tencent Cloud infrastructure, Hunyuan is effectively mandatory for evaluating — the integration depth is significant. For builders outside that ecosystem, HunyuanVideo is the most compelling open-weight offering: it was rated among the top 3 open-source video generation models in multiple independent evaluations in Q1 2025.

Baidu ERNIE — RAG-optimized, search integration

Baidu's ERNIE series has the longest lineage of any Chinese LLM family (ERNIE 1.0 launched in 2019) and in 2026 has found its most defensible position: deep integration with Baidu Search and Baidu Cloud for RAG-intensive workloads. ERNIE 4.0 Turbo (2024) is competitive with GPT-4o on Chinese-language tasks and specifically excels at tasks requiring knowledge retrieval and synthesis — the core use case for Baidu's search-adjacent products.

Key model: ERNIE 4.0 Turbo, ERNIE Speed (cost-optimized), ERNIE Character (persona fine-tuning) License: Proprietary API; Wenxin Workshop enterprise license for deployment API Access: Baidu AI Cloud / Wenxin Workshop (qianfan.cloud.baidu.com); international accounts supported with some regional restrictions Best for: Chinese knowledge-intensive RAG, Baidu Search integration, enterprise document Q&A, long-context Chinese text summarization, companies with Baidu Cloud infrastructure GitHub / HF: huggingface.co/baidu (limited); primary surface is Wenxin Workshop API docs

Baidu has invested heavily in making ERNIE accessible to enterprise builders through their Qianfan platform — it has the most comprehensive Chinese-language enterprise support tier of any Chinese AI lab, including guaranteed SLA, private deployment options, and MLOPS tooling. The tradeoff is that it's the least open of the six companies here: ERNIE model weights are not publicly released.

Zhipu AI / GLM — coding and agents, open weights

Zhipu AI is the lab most underestimated by English-language coverage and most valued by Chinese developers. Their GLM-4 series (released early 2024, updated through 2025) is consistently rated as the best Chinese-language coding model by Chinese developer communities, and their open-weights philosophy (GLM-4-9B, GLM-4V-9B available on HuggingFace under a permissive license) makes it uniquely builder-accessible among the non-Apache-2.0 options.

Key model: GLM-4 (flagship), GLM-4-9B (open weights), GLM-4V (vision), CogVideoX (video generation), CogView (image generation) License: GLM-4-9B — model license (permissive for most commercial uses, similar to Llama license); GLM-4 API — proprietary API Access: Zhipu AI BigModel (bigmodel.cn / open.bigmodel.cn); international access available; documentation in both Chinese and English Best for: Chinese-language code generation and documentation, agent workflows, multimodal (CogVideoX is a top-tier video generation model), self-hosted small model with strong Chinese reasoning GitHub / HF: huggingface.co/THUDM (full model weights including GLM-4-9B), github.com/THUDM

CogVideoX-5B (released August 2024) became one of the most downloaded open-source video generation models on HuggingFace within weeks of release. For Chinese developer communities, GLM-4 is often preferred over Qwen for code generation tasks because of its training data composition — more Chinese technical documentation and open-source Chinese codebases. GLM-4-9B in particular is a strong choice for self-hosted deployments where you need a bilingual model under 10B parameters.

Company comparison table

How to evaluate a Chinese AI company for builder use

Before integrating any Chinese AI API into a production workflow, run three quick tests:

The Reproducibility Test: Can you reproduce the benchmark claims? For models on HuggingFace, run a quick eval on a task you care about using their provided evaluation scripts. For API-only models, check if the lab has published a technical report with specific benchmark methodology. ERNIE, Baichuan, and Hunyuan-T1 have been the least reproducible in third-party evaluations; DeepSeek, Qwen, Yi, and GLM have been most reproducible.

The API Availability Test: Sign up for the API with an international account and make 10 test API calls across 3 days. Check: Does signup work without a Chinese phone number? Does billing work with an international card? Are there hard quota limits that would block a real workload? MiniMax, 01.AI (via HF), and Zhipu AI BigModel pass this test most consistently for global builders. Baichuan enterprise requires direct contact. ERNIE/Baidu has the most friction for international accounts.

The License Test: Read the license file, not just the label. "Open source" in Chinese AI press releases has been applied to models with varying degrees of actual openness. Apache 2.0 (Yi, Qwen3) means fully commercial. MIT (DeepSeek) is equivalent. Custom model licenses (GLM, Baichuan) require reading the specific terms. Proprietary API-only models (ERNIE, Hunyuan-T1) require a commercial agreement.

Monthly monitoring checklist per company

Run this checklist once a month to stay current without doomscrolling:

MiniMax: Check minimax.io/news for model updates; check HunyuanVideo HuggingFace page for download stats (proxy for community adoption); monitor @HailuoAI on X for international product launches.

Baichuan: Low-cadence for most builders. Check baichuan-ai.com once a quarter unless you're in healthcare or legal verticals. Monitor Chinese tech news (36Kr, KR Asia) for enterprise contract announcements.

01.AI: Check huggingface.co/01-ai for new model releases; activity has been lower since mid-2024 but the Yi-1.5 series remains production-ready. Check 01.ai for API updates.

Tencent Hunyuan: Check huggingface.co/tencent for open-weight releases (HunyuanVideo, Hunyuan3D updates are most active). For API/cloud updates, check Tencent Cloud AI product pages.

Baidu ERNIE: Check Wenxin Workshop release notes (qianfan.cloud.baidu.com/doc) for API pricing and context window updates. ERNIE makes frequent model upgrades that affect API behavior. Monitor for pricing tier changes — they've adjusted free tier quotas multiple times in 2025.

Zhipu AI GLM: Most active of the six on HuggingFace — check huggingface.co/THUDM for new model weights. CogVideoX updates are frequent. Check open.bigmodel.cn for API changelog. Follow @zhipuai_cn on X for major release announcements.

FAQ

Is MiniMax available outside China? Yes. MiniMax operates an international API surface at minimax.io with English documentation and international billing (Visa/Mastercard). The Hailuo AI brand handles international video generation. As of May 2026, signups without a Chinese phone number are supported. API quotas are more generous on the international tier than many China-based providers.

What is the Yi model? Yi is the foundation model series from 01.AI, co-founded by Kai-Fu Lee. Yi-34B (2023) was one of the first Chinese LLMs to perform competitively on English-language reasoning benchmarks alongside Chinese. The series was updated to Yi-1.5 in May 2024 with improved context handling and instruction following. All Yi models are Apache 2.0 licensed, making them among the most permissively licensed Chinese foundation models available.

GLM vs Qwen — which is better for coding? It depends on the language. For English-language code generation, Qwen2.5-Coder-32B is generally rated higher. For Chinese-language code generation (Chinese variable names, Chinese comments, Chinese documentation), GLM-4 is often preferred by Chinese developer communities. For agent workflows with tool calling and Chinese-language planning, GLM-4 has better benchmarks on Chinese-specific agent tasks. For mixed bilingual codebases, this is worth benchmarking on your specific workload.

Can I use Baichuan commercially? Baichuan AI's earlier open models (on HuggingFace) have licenses allowing commercial use with attribution. Baichuan-4, the current flagship, is proprietary and requires an enterprise license agreement. For most international builders, commercial Baichuan use means going through their API platform with a paid tier — direct contact with their sales team is required for enterprise SLA commitments. For healthcare or legal verticals in China, Baichuan-M1-235B commercial deployment requires a formal contract.

Which Chinese AI company is best for building agents? Zhipu AI GLM-4 has the strongest Chinese-language agent capabilities based on AgentBench scores and community usage. For English-first agent workflows, DeepSeek-V3 and Qwen3 are ahead. For tool-calling and function-calling APIs, both GLM-4 and Qwen3 have mature function calling implementations. MiniMax-Text-01 is worth evaluating for agents that need long context windows (1M tokens) with multimodal input.

How does Tencent Hunyuan compare to other Chinese multimodal models? Hunyuan's text model (T1) is competitive with top-tier Chinese models but not the market leader. Where Hunyuan genuinely leads is HunyuanVideo: as of Q1 2025, it was rated among the top 3 open-source video generation models globally. For builders specifically needing video generation with open weights and a commercial license, HunyuanVideo (Apache 2.0) is the go-to Chinese option.

What happened to Baidu ERNIE 4.0? ERNIE 4.0 (released October 2023) was Baidu's response to GPT-4 and showed strong performance on Chinese benchmarks. ERNIE 4.0 Turbo (2024) improved speed and cost efficiency. The ERNIE family is now part of Baidu's Qianfan enterprise AI platform. Model weights have never been released publicly — all access is API-only through Wenxin Workshop. Baidu's focus is on deep Baidu Search integration rather than competing for open-source mindshare.

Is Zhipu AI GLM open source? Partially. GLM-4-9B and GLM-4V-9B model weights are available on HuggingFace (huggingface.co/THUDM) under a permissive model license suitable for most commercial applications. The larger GLM-4 flagship is API-only. CogVideoX-5B (video generation) is available on HuggingFace under a permissive license. The GLM training code and fine-tuning recipes are publicly available on GitHub (github.com/THUDM/GLM-4).

Evidence and source links

• MiniMax API: minimax.io — international API surface with English docs
• Baichuan AI: baichuan-ai.com — enterprise model page
• 01.AI Yi models: huggingface.co/01-ai — full model weights, Apache 2.0 license files
• Tencent HunyuanVideo: huggingface.co/tencent/HunyuanVideo — Apache 2.0 open weights
• Zhipu AI THUDM: huggingface.co/THUDM — GLM-4 series weights and code
• Baichuan HuggingFace: huggingface.co/baichuan-inc — earlier open model versions
• Baidu Wenxin Workshop: qianfan.cloud.baidu.com — ERNIE API documentation and pricing

Companion pages in this cluster

• Foundation model landscape: /en/china-ai-foundation-models — full model family comparison with benchmark data
• Open source model licensing: /en/china-ai-open-source-models — Apache 2.0, MIT, custom license breakdown
• China AI models watchlist: /en/china-ai-models-list — current tracking list with release monitoring
• API access guide: /en/china-ai-api-access-guide — how to access Chinese AI APIs globally
• China AI in 2026 overview: /en/china-ai-in-2026 — complete landscape reference

Related reading

• China AI Updates in English: What Builders Should Watch Each Month
• Top China-Built AI Models to Watch in 2026: DeepSeek, Qwen, Kimi & More
• How to Track China AI in English Without Doomscrolling
• Best English Sources for China AI Industry Updates (2026 Guide)

RadarAI helps builders track AI updates, compare source-backed signals, and decide which changes are worth acting on.