China's Foundation Models

Learning Objectives

• Identify the major Chinese AI companies and their leading foundation models
• Explain why DeepSeek's efficiency breakthrough mattered so much to the global AI industry
• Explain why China made open source a national strategy, and what that buys a country constrained on hardware
• Assess how close open-weights models have come to the proprietary frontier — and where they still fall short
• Understand the data privacy implications of using Chinese-hosted AI APIs vs. running Chinese open-source models locally

The Data Privacy Warning You Need to Read First

Before exploring Chinese AI models, there is an important distinction you must understand.

Why Chinese AI Models Matter

Until late 2024, the assumption in the global AI industry was that frontier AI required American labs, American chips, American talent, and hundreds of millions of dollars in training compute. DeepSeek changed that assumption overnight.

China has an enormous AI ecosystem built across university research (Tsinghua, Peking University, Shanghai Jiao Tong), tech giants (Alibaba, Baidu, Tencent, ByteDance, Huawei), and well-funded startups. The models coming out of this ecosystem are not derivatives of US models — they are independently developed, often built under significant hardware constraints due to US chip export controls.

The result is a set of models that are competitive with US frontier models on key benchmarks, often with dramatically higher cost efficiency.

The Hardware Story Has Flipped — Inside China

The hardware-constraint framing also needs an update. As of May 2026, NVIDIA CEO Jensen Huang publicly acknowledged that NVIDIA has "largely conceded" the Chinese AI accelerator market — the company's share inside China is now near zero. After more than a year of US export controls and a stalemate where H200 customs clearance letters have sat unprocessed, Beijing's homegrown-stack push has paid off: Alibaba, ByteDance, and Tencent are all running production AI workloads on Huawei's Ascend silicon, and DeepSeek's V4 release in April 2026 was optimized for Ascend before CUDA. Huawei now expects roughly $12 billion in AI accelerator revenue in 2026 — up from $7.5 billion in 2025 — on the strength of those three customers alone. Inside China, the chip-export constraint that motivated DeepSeek's original efficiency push has effectively been replaced by a parallel domestic supply chain.

Open Source Is Now Chinese National Strategy

The most important development in this lesson is not a model. On July 17, 2026, Xi Jinping opened the World AI Conference in Shanghai and said China would seize a "rare historic opportunity" by "encouraging open source, openness, collaboration and sharing." It was the first time China's head of state framed open weights as national industrial policy rather than a choice individual labs were making.

Xi argued that AI development "should not be a solo performance by any single country but rather a symphony of global cooperation," and criticized broad national-security claims that let powerful nations restrict AI access for others — an unmistakable reference to US export controls. China also pledged 5,000 AI research and training placements for developing countries over the next five years.

Read the timing. Moonshot AI shipped Kimi K3 the day before — the largest open-weights model any Chinese lab has produced, and the first to beat a leading US proprietary flagship on most published coding benchmarks. The policy statement and the proof point landed within twenty-four hours of each other.

DeepSeek — The Efficiency Revolution

DeepSeek AI is a research lab founded by the Chinese hedge fund High-Flyer Capital Management, based in Hangzhou. It released a series of models beginning in 2024 that stunned the global AI community.

The DeepSeek Market Shock

On January 27, 2025, DeepSeek R1 triggered the largest single-day market loss in US stock market history. NVIDIA lost $589 billion in market value in a single day — more than double the previous record. The Nasdaq fell 3.1%, the S&P 500 dropped 1.5%, and approximately $1 trillion was wiped from US markets in total.

The trigger: DeepSeek demonstrated ChatGPT-level capabilities at a claimed training cost of $5.6 million, versus hundreds of millions for US competitors. The market question was existential: if frontier AI could be built cheaply, was the massive investment in AI infrastructure justified?

That figure is now contested. In a joint advisory published on September 8, 2026, the US National Security Agency, Federal Bureau of Investigation and Cybersecurity and Infrastructure Security Agency said that figure leaves out the cost of data DeepSeek obtained by distilling US frontier models, and named the lab as one of six Chinese companies running what the agencies call industrial-scale distillation campaigns since late 2024. China's foreign ministry calls the accusations groundless. Note what this does and does not change: the engineering efficiencies DeepSeek published are independently reproducible and were never in dispute, and the market reaction happened regardless. What is in dispute is whether the headline number represents the full cost of building the model — which matters, because the number is what made the shock a story about economics rather than about capability.

Two days later a vendor published its own logs. On September 10, 2026 Anthropic released a threat intelligence report attributing distillation campaigns, with what it calls high confidence, to Alibaba, Moonshot AI, DeepSeek, Zhipu, Xiaomi, SenseTime and MiniMax. The overlap with the government list is five of seven — Xiaomi and SenseTime are new, and StepFun does not appear — so the two accounts are not interchangeable and neither supersedes the other.

What the report adds is specificity the advisory could not provide. Anthropic says Alibaba ran the largest campaign it has measured: more than 151 million exchanges between May and July 2026, peaking near three million a day across thousands of fraudulent accounts, harvesting the chain-of-thought reasoning behind Opus-class models and using it to train Qwen 3.5, 3.6 and 3.7. Moonshot is credited with more than 23 million exchanges over the same window and Zhipu with more than 3.4 million over seventeen days.

The most consequential claim is not about theft at all. Anthropic says Moonshot and DeepSeek silently routed their own customers' requests to Claude and kept the transcripts — users believed they were talking to Kimi or DeepSeek and received Claude's answers instead. Those relayed sessions, it says, carried names, contact details, corporate data and live credentials belonging to people who never chose to expose them to a third party. If accurate, that is a privacy failure by the labs against their own paying users, which is a separate matter from whatever the harvested reasoning was worth.

The answer, ultimately, was yes. Markets fully recovered. NVIDIA became the first company to reach a $5 trillion valuation by October 2025. Cheaper AI increased demand for AI compute rather than reducing it. But the DeepSeek shock permanently changed assumptions about the relationship between compute spending and AI capability.

DeepSeek V4 — Current Flagship (April 2026)

On April 24, 2026, DeepSeek released V4-Pro (1.6 trillion total parameters, 49 billion active, mixture-of-experts) and V4-Flash (284 billion total / 13 billion active), both MIT-licensed and shipping with a 1 million-token context window — the first DeepSeek models to match Claude and Gemini on context length. V4-Pro was the largest open-weights model at the time, since passed by Kimi K3 (2.8 trillion) and Qwen3.8 Max (2.4 trillion), and it remains the largest under a genuinely permissive license. The accompanying paper claims V4-Pro uses approximately 27% of V3.2's FLOPs and 10% of the KV cache at 1 million-token context — meaningful both for training cost and on-device inference.

API pricing moved twice in one month, and in opposite directions. The 0813 build cut V4-Pro by about three quarters when it reached general availability on August 12, to roughly 44 cents per million input tokens and 87 cents per million output. Four days later DeepSeek reversed that and more. From August 16, 2026 the API bills at peak and off-peak rates, with peak hours running from 1 to 4 and from 6 to 10 UTC and off-peak charged at half:

• V4-Pro: 66 cents in and one dollar ninety-eight out off-peak; one dollar thirty-two in and three dollars ninety-six out at peak
• V4-Flash: 22 cents in and 66 cents out off-peak; 44 cents in and one dollar thirty-two out at peak

Cached input rose furthest — roughly 1,100 percent on V4-Pro. The lab had warned in its own documentation that a significant increase was coming without naming a date; this is that increase. The strategic reading matters more than the numbers: DeepSeek is no longer competing on being cheapest. At peak, V4-Pro now costs more per output token than GPT-5.6 Luna, which OpenAI cut to 20 cents in and one dollar twenty out on July 30. The cheap-Chinese-model assumption that shaped 2025 no longer holds automatically in either direction.

Both are available via DeepSeek's API, on Hugging Face for self-hosting, and through third-party providers (Together.ai, Fireworks AI, Groq). V4-Flash scores 50 on the Artificial Analysis Intelligence Index, third among open-weights models — the clearest single anchor for where the open Chinese stack sits against the proprietary frontier.

DeepSeek R2 never shipped as a standalone model. The reasoning engine it was meant to carry landed inside V4 instead, so the lab's reasoning lineage runs R1 to R1-0528 to V4 rather than through a separate R-series flagship.

First Outside Venture Round — AGI Mandate (June 2026)

DeepSeek closed its first outside venture round in June 2026, raising approximately 50 billion yuan (about $7.4 billion) — the first external capital the lab had ever taken. Commercial investors were led by Tencent, at 10 billion yuan, and battery maker CATL, at 5 billion yuan, with NetEase and JD.com also participating. Founder Liang Wenfeng contributed another 20 billion yuan from his own holdings.

The terms are unusual. Commercial backers accepted five-year lockups and no voting rights, and the only investor granted governance rights and a direct stake was Beijing's National Artificial Intelligence Industry Investment Fund, China's state-backed strategic AI vehicle (distinct from the chip-focused "Big Fund"). Liang had not previously sought outside capital, and the round is framed as a way to offer employee equity and retain talent against intensifying domestic competition — without loosening his grip on direction. He has said publicly that the lab will keep developing open-source models and pursue artificial general intelligence as its core goal, resisting the usual pressure to chase near-term commercialization.

DeepSeek R1 — The Open-Source Reasoning Breakthrough

DeepSeek R1 was the first open-source reasoning model to match OpenAI's o1 on challenging benchmarks (AIME math, GPQA science).

Reasoning models spend more compute at inference time "thinking through" a problem before responding — similar to how a human might work through a math proof step by step rather than guessing immediately. DeepSeek R1 demonstrated that this "extended thinking" capability — previously exclusive to OpenAI's o1 — could be replicated and open-sourced.

R1-0528 (May 2025) was a major revision that improved logic and programming benchmarks, and added JSON output and function-calling capabilities. Distilled versions (1.5 billion, 7 billion, 14 billion, 32 billion, 70 billion parameters) allow the reasoning capability to run on modest hardware.

Previous Generation — V3.2 and V3.2-Speciale

V3.2 was DeepSeek's flagship before V4 shipped in April 2026. It is still widely deployed and remains available via API for cost-sensitive workloads, but new builds should target V4.

DeepSeek V3.2 is a 671-billion-parameter mixture-of-experts model released under the MIT license. What made it remarkable in early 2025:

• Training cost: Approximately $5.9 million in compute — compared to estimates of $100 million+ for comparable US models
• Performance: At release it ranked as the #2 most intelligent open-weight model on the Artificial Analysis Intelligence Index, ahead of Grok 4 and Claude Sonnet 4.5 (Thinking) — a position it has since given up to newer open-weights releases, including DeepSeek's own V4 line
• Coding: 40%+ improvement on SWE-bench Verified over V3.1
• MIT license: Completely free for commercial and research use

V3.2-Speciale (December 2025) pushed even further — a high-compute variant that won a gold medal at the 2025 International Mathematical Olympiad (35/42 points), placed 10th at the International Olympiad in Informatics, and scored 96.0% on AIME (vs. GPT-5-High's 94.6%). The Speciale API was only available until December 15, 2025, due to extreme compute costs.

DeepSeek Bans and Restrictions

DeepSeek's success has been accompanied by significant regulatory pushback:

• Italy blocked DeepSeek from app stores (January 2025) over GDPR failures
• Banned on government devices in South Korea, Australia, and Taiwan
• US restrictions: banned by NASA, the US Navy, the Pentagon, and on Texas state government devices; restricted in US House offices

These bans target the hosted API (where data flows to Chinese servers), not the open-source weights that can be run locally.

Alibaba — Qwen Series

Alibaba DAMO Academy has built one of the broadest open-source model portfolios in the world with the Qwen (通义千问, Tongyi Qianwen) series.

Qwen3.5

The Qwen3.5 family (February–March 2026) was the family's early-2026 open-weight generation, spanning from 0.8 billion to 397 billion parameters.

The flagship is Qwen3.5-397 billion-A17 billion — a 397 billion total parameter MoE model with only 17 billion active per forward pass. It uses a novel Gated DeltaNet + MoE architecture that alternates linear and full attention in a 3:1 ratio for efficiency.

Key characteristics:

• 262K native context window, extensible to 1 million tokens
• 100+ languages — broader multilingual coverage than most US models, with particular strength in Chinese, Japanese, Korean, Arabic, and less-resourced languages
• The 9 billion model matches or surpasses GPT-OSS-120 billion (13-times its size) on GPQA Diamond and MMMU-Pro benchmarks — exceptional efficiency
• The 122 billion-A10 billion variant scores 72.2 on tool use benchmarks (vs. GPT-5 mini's 55.5)
• Released under Apache 2.0 license

The Qwen3.5 family includes medium models (27 billion, 35 billion, 122 billion) and small models (0.8 billion, 2 billion, 4 billion, 9 billion), covering everything from on-device deployment to large-scale server inference.

Qwen3.6, 3.7, 3.8, and licensing by tier

Alibaba iterated fast after Qwen3.5, and in the process changed its release strategy. Qwen3.6 (April 2026, Apache 2.0) continued the open-weight line — a 27-billion-parameter dense model plus a 35-billion-A3-billion mixture-of-experts model, with the compact 27 billion matching or beating the far larger Qwen3.5 on agentic coding. The mid-tier flagships went proprietary: Qwen3.7 Max (a text reasoning and agentic model, May 2026) and Qwen3.7 Plus (a multimodal agent, generally available since June 2026) are API-only, with a 1 million token context window.

Qwen3.8 Max landed in August 2026 and filled in every blank the July preview had left. It is a mixture-of-experts model with 2.4 trillion total parameters and roughly 95 billion active per request, 262,144 tokens of native context, and a published benchmark sheet: 86.6 on Terminal-Bench 2.1 — ahead of Claude Opus 4.8 and Claude Fable 5 at 84.6, behind GPT-5.6 Sol at 88.8 — against a much weaker 67.7 on SWE-bench Pro, where Fable 5 scores 80.0. That spread is the honest summary: Qwen3.8 Max is genuinely competitive on agentic and multimodal work while still trailing the US frontier on hard software engineering. The hosted API runs $2 per million input tokens and $6 per million output tokens.

On August 12, 2026, roughly nine days after the announcement, Alibaba published the Qwen3.8 Max weights. At 2.4 trillion parameters it is the second-largest open-weight release to date, behind Kimi K3's 2.8 trillion. So the picture is not two clean tracks but licensing by tier: small and mid-sized models under Apache 2.0, the flagship downloadable under custom terms, and the intermediate Max and Plus models still API-only.

QwQ-32 billion

QwQ-32 billion is Alibaba's reasoning-specialized model at 32 billion parameters. It is designed for mathematical reasoning and step-by-step problem solving, competing directly with DeepSeek R1's smaller distilled variants.

Qwen Wins the Apple Slot (July 2026)

On July 15, 2026, China's Cyberspace Administration cleared Apple Intelligence for release in China, with Qwen as the model handling text and image understanding and generation across iOS, iPadOS, macOS, and visionOS for Chinese users. Baidu contributes a smaller model alongside it. Apple evaluated Baidu, DeepSeek, and ByteDance before settling on Alibaba.

This is the most consequential distribution win any Chinese model has achieved. Elsewhere in the world, the rebuilt Siri's cloud requests run on a custom Google Gemini model — but Chinese regulation requires a domestically approved model, so inside China the brain of Apple's AI is Qwen. That places Alibaba's models in front of one of the largest premium device populations on earth, in Apple's biggest market outside the United States.

Moonshot AI — Kimi

Moonshot AI is a Beijing-based startup founded in 2023, focused on long-context applications.

Kimi K2.5 (January 2026) was a major leap from the earlier K2:

• 1 trillion parameters MoE architecture with 32 billion active per forward pass
• 256K context window (doubled from K2's 128K)
• Trained on 15 trillion mixed visual and text tokens — natively multimodal
• Outperforms Gemini 3 Pro on SWE-Bench Verified and beats GPT 5.2 on SWE-Bench Multilingual
• Beats GPT 5.2 and Claude Opus 4.5 on VideoMMU (video understanding)

Moonshot also released Kimi Code — an open-source coding tool with integrations for terminals, VS Code, Cursor, and Zed.

Kimi K2.6 + $2 Billion Round (May 2026)

On May 7, 2026, Moonshot AI shipped Kimi K2.6 alongside a $2 billion funding round at a $20 billion valuation. The round was led by Meituan's Long-Z Investments arm, with Tsinghua Capital, China Mobile, and CPE Yuanfeng participating. Moonshot reported $200 million in annualized recurring revenue in April 2026 — the kind of consumption disclosure most Chinese labs do not make publicly.

The Moonshot round landed in the same stretch of weeks that DeepSeek was assembling its own first outside round (covered above), which closed the following month. Together, the two raises signalled that open-weights labs out of China are emerging as the primary cost-pressure on US frontier vendors — DeepSeek leading on raw scale and government-aligned capital, Moonshot on direct-API consumption volume and OpenRouter ranking.

K2.6 retains K2.5's 1 trillion parameter MoE architecture (32 billion active per token), extends the context window to 262K tokens, adds native INT4 quantization, and introduces an Agent Swarm system that scales to 300 sub-agents and 4,000 coordinated steps in 12-hour autonomous coding sessions. Open-weights landed on Hugging Face on April 20, 2026 under a Modified MIT License. See Kimi K2.6 (Moonshot AI) for the full architecture and Cursor's deliberate K2.5 + RL versus K2.6 generation-skip decision.

Kimi K3 — The Open Frontier Arrives (July 2026)

On July 16, 2026, Moonshot shipped Kimi K3 and K2.6 became the prior generation. K3 is the most consequential open-weights release to date, and it is worth being precise about why.

K3 is a 2.8 trillion parameter mixture-of-experts (MoE) model that activates just 16 of 896 experts per token — far sparser than the K2 family — with a 1 million token context window and native vision. It is built on two techniques Moonshot developed in-house: Kimi Delta Attention, the basis for the million-token context, and Attention Residuals, which retrieves representations selectively across model depth rather than accumulating them uniformly.

The headline is not the size. It is that K3 beats Claude Opus 4.8 on most of the published coding and agentic suite — most dramatically on FrontierSWE, 81.2 against 66.7. For roughly two years, the open-weights argument carried an implicit concession: open models were cheaper and more private, but a generation behind. K3 is the strongest evidence yet that the concession is narrowing.

Two things keep this from being a clean victory lap, and both matter:

• K3 does not beat the current frontier. Moonshot says so itself — K3 "still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol," and names user experience, not capability, as the remaining gap. On HLE-Full, a hard reasoning benchmark, K3 scores 43.5 against Fable 5's 53.3.
• Open weights are not free to run. A 2.8 trillion parameter model is far beyond consumer hardware — the published checkpoint spans 96 shards and well over a terabyte. Most teams will consume K3 through a hosted endpoint anyway, which puts the data-residency question right back where it started.

K3's pricing signals the shift most clearly. At $15 per million output tokens, Moonshot is no longer undercutting the frontier the way earlier Chinese open models did. It is pricing like a frontier lab.

Kimi K3

Moonshot AI

Open Source

Strengths

2.8 trillion MoE (16 of 896 experts active); 1 million context; native vision; beats Claude Opus 4.8 on most coding benchmarks; largest open-weights model from a Chinese lab

Context Window

1 million tokens

Pricing

Free tier on kimi.com; API at 30 cents per million cached input and $15 per million output; weights public since July 27, 2026 under the Kimi K3 License

Moonshot AI →

Baidu — ERNIE 5.0

Baidu is China's leading search engine company, analogous to Google. Its ERNIE (文心一言, Wenxin Yiyan) series has reached its fifth generation.

ERNIE 5.0 (November 2025, announced at Baidu World 2025) is a fundamental upgrade:

• 2.4 trillion parameters — a unified multimodal model integrating text, image, video, and audio in a single autoregressive framework
• Comparable to Gemini-2.5-Pro and GPT-5-High on 40+ authoritative benchmarks
• Dominant performance on Chinese-language benchmarks
• Real-time search grounding via Baidu's search integration
• Closed source; available via Baidu Cloud API

Baidu is also developing its own AI chips: Kunlunxin M100 (optimized for inference, releasing early 2026) and Kunlunxin M300 (for training/inference of ultra-large models, following in early 2027).

Best for: Applications where Chinese-language fluency is paramount and integration with Chinese-language web search adds value.

ERNIE 5.0

Baidu

Closed

Strengths

2.4 trillion params; unified multimodal (text/image/video/audio); Chinese-language leader; Baidu search integration

Context Window

Not disclosed

Pricing

Baidu Cloud API

Baidu →

Zhipu AI — GLM-5

Zhipu AI is a spinout from Tsinghua University's AI lab. In January 2026, Zhipu made history as the first publicly listed Chinese AI foundation model company, raising approximately $558 million in a Hong Kong IPO.

GLM-5 (February 2026) is a generational leap from GLM-4.5:

• 744 billion total parameters MoE, with 256 experts and 8 activated per token (~44 billion active)
• Uses DeepSeek Sparse Attention (DSA) for efficiency
• 200K context window
• MIT license (fully open source)
• Claims to surpass Gemini 3 Pro on coding and agentic performance
• Trained entirely on Huawei Ascend chips using MindSpore framework — zero NVIDIA dependency

GLM-5 is significant not just for its performance, but because it proves that frontier-class models can be trained without any NVIDIA hardware — a milestone for Chinese AI independence.

GLM-5-Turbo (March 2026) is optimized specifically for automated agent workflows.

GLM 5.2 (June 2026) extended the line with a 1 million token context window and a coding-first focus, rolling out across Zhipu's GLM Coding Plan tiers and pitched as a permissively licensed alternative to closed frontier coding assistants for the Asia-Pacific market. The standalone API and MIT-licensed open weights both shipped, and the benchmarks are now public: 62.1 on SWE-bench Pro (ahead of GPT-5.5 at 58.6) and 74.4 percent on FrontierSWE, a near-tie with Claude Opus 4.8.

GLM-5.3 (August 14, 2026) is the current flagship, and it inverts the usual pattern in two ways. It reuses GLM 5.2's base model, taking every gain from scaled-up post-training rather than a new pre-training run — lifting CyberGym from 77.2 to 84.5 percent and DeepSWE from 46.2 to 66.9. And Z.ai withheld the open weights for roughly two weeks of safety evaluation, saying cyber capability grew faster than it expected as training scaled, with the model planning complete exploitation chains rather than isolated bugs.

Both checkpoints have now shipped, and the split between them is the useful part. GLM-5.3-Flash is plain MIT — ungated, with no revenue threshold and no geographic carve-out. The flagship, published on August 28, carries its own GLM-5.3 License instead: an MIT-shaped grant plus one condition, requiring a Z.ai security review before commercial use by any model-as-a-service operator whose revenue passes 10 billion dollars over twelve consecutive months. Almost no one meets that bar, so the practical freedom is close to identical — but the terms are genuinely different, and they differ within a single generation released two weeks apart. Read the license attached to the specific checkpoint you intend to run, not the one on its sibling.

GLM-5

Zhipu AI

Open Source

Strengths

744 billion MoE (44 billion active); 200K context; MIT license; trained entirely on Huawei Ascend; competitive with Gemini 3 Pro

Context Window

200K tokens

Pricing

Free (self-hosted); API available

Zhipu AI →

Tencent — Hunyuan

Tencent gets less Western coverage than the labs above and matters at least as much, because Hunyuan models ship inside WeChat, QQ and Tencent Cloud — distribution no Western lab can approach inside China.

Hy4 preview (August 28, 2026) is the current flagship: a mixture-of-experts model with 770 billion total parameters, roughly 49 billion active per token, and a context window past one million tokens. Tencent reports an internal blind evaluation, 163 experts across 203 engineering tasks, scoring it 2.99 out of 4.00 against 2.92 for GLM-5.3 and 2.94 for Kimi K3. That is a vendor-run test on a vendor-authored benchmark, so read it as a claim of rough parity with the open frontier rather than a measured win. Weights are published on Hugging Face, ModelScope, GitCode and CNB.

The license is where this release earns its place on this page, because Tencent moved in the opposite direction from every other lab here. Hy4's LICENSE file is plain Apache 2.0 — no territorial carve-out, no revenue trigger, no user threshold. Earlier Hunyuan releases ran under the Tencent Hunyuan Community License, which defines its "Territory" as the world excluding the European Union, the United Kingdom and South Korea, and requires a separate agreement from Tencent for any product above 100 million monthly active users. Under those terms a European or British developer had no license to use the weights at all.

So Tencent is the clearest case of the rule this page keeps arriving at, running the other way. Alibaba, Moonshot and Z.ai each tightened terms on a newer or larger checkpoint; Tencent loosened them on its largest model while leaving a smaller sibling restricted. The license belongs to the checkpoint, not to the vendor, the generation, or the parameter count.

Hy4 preview

Tencent

Open Source

Strengths

770 billion total / 49 billion active mixture-of-experts; context past one million tokens; plain Apache 2.0 with no territorial, revenue, or user limits

Context Window

1 million+ tokens

Pricing

Free (self-hosted); API available via Tencent Cloud

Tencent →

ByteDance — Doubao

ByteDance (the company behind TikTok) has quietly built the most-used AI chatbot in China with Doubao (豆包).

Doubao 2.0 / Seed 2.0 (February 2026) comes in four variants — Pro, Lite, Mini, and Code — with the Pro variant delivering:

• 98.3% on AIME 2025 (math), a 3020 Codeforces rating (competitive programming), and 89.5 on VideoMME (video understanding)
• Performance matching GPT 5.2 and Gemini 3 Pro at roughly 1/10th the cost
• Doubao has exceeded 100 million daily active users and 155 million weekly active users

ByteDance's strategy is distinctive: rather than competing for frontier research prestige, it leverages its massive TikTok/Douyin distribution network to put AI in the hands of hundreds of millions of users at extremely low prices.

Doubao 2.0 Pro

ByteDance

Closed

Strengths

98.3% AIME; 100 million+ DAU; GPT-5.2-competitive at 1/10th cost; most-used AI chatbot in China

Context Window

Not disclosed

Pricing

API available; consumer app free

ByteDance →

MiniMax — Consumer AI Powerhouse

MiniMax is a Shanghai-based AI company that IPO'd in Hong Kong in January 2026, raising approximately $620 million — with its stock surging 109% on debut. Backed by Alibaba, Tencent, and Abu Dhabi Investment Authority.

MiniMax M2.7 (March 2026) is its latest model. MiniMax is known for consumer-facing AI products including AI companion apps and creative tools, with over 100 million users.

MiniMax and Zhipu AI were among the first of China's "Six Tigers" — the six leading Chinese AI startups (Zhipu, Moonshot, MiniMax, Baichuan, StepFun, 01.AI) — to go public. Notably, 01.AI stopped pre-training large models in March 2025, pivoting to selling business solutions using DeepSeek's models — a significant strategic retreat that highlighted the brutal economics of competing in foundation models.

MiniMax M2.7

MiniMax

Closed

Strengths

Consumer AI focus; Hong Kong IPO ($620 million raised); 100 million+ users; backed by Alibaba/Tencent

Context Window

Not disclosed

Pricing

API available; consumer apps free

MiniMax →

Ant Group — InclusionAI / Ling Series

Ant Group is the financial-services affiliate of Alibaba (operator of Alipay) and one of the largest fintech companies in the world. Ant's AGI initiative, InclusionAI, sits inside the firm and ships open-weights frontier models alongside its applied financial-AI work — putting Ant in a different posture than DeepSeek (a hedge-fund spin-out) or Alibaba's Qwen team (a hyperscaler model lab).

In early May 2026, InclusionAI published Ling-2.6 — including a 1 trillion parameter flagship variant — on Hugging Face under the MIT license. The architecture is a hybrid of Multi-head Latent Attention and Linear Attention with a 262,144-token context window. Headline benchmark: 72.2 on SWE-bench Verified, among the strongest scores any open-weights model has posted on a coding evaluation. Inference requires tensor parallelism across 8 GPUs. A companion hosted-only sibling, Ring 2.6 at the same trillion-parameter scale, surfaced on OpenRouter at the same time.

Ling-2.6 strengthens the broader pattern from this module: Chinese labs are not just shipping models that compete on benchmarks — they're shipping them under permissive licenses that anyone can self-host and inspect. That's a different distribution strategy from US frontier labs, where flagship weights are closed and only research-grade or older-generation models go open. For developers and security-conscious enterprises that need to run models inside their own perimeter, the open-weights Chinese stack is increasingly the only option at the trillion-parameter scale.

Ling-2.6 (1T variant)

Ant Group

Closed

Strengths

1 trillion parameter open weights under MIT license; 72.2 on SWE-bench Verified; 262,144-token context; hybrid MLA + Linear Attention; from Ant Group's InclusionAI lab

Context Window

262,144 tokens

Pricing

Free open weights; tensor parallelism across 8 GPUs required for inference

Ant Group →

The US Chip Export Controls Context

Understanding Chinese AI development requires understanding the constraint it operates under: US export controls.

Since 2022, the US Bureau of Industry and Security (BIS) has progressively restricted exports of advanced AI chips to China. The original H100 and H200 remain banned. In December 2025, the Trump administration partially reversed course by allowing H200 exports to China with a 25% surcharge (up to 80,000 chips) — better to sell older chips with surcharges than have China build its own. Blackwell-class chips (B100, B200, GB200) and next-generation Rubin chips remain fully restricted for 18–24 months after domestic launch.

China has responded on multiple fronts:

Huawei Ascend chips are rapidly advancing:

• Ascend 910C: 600,000 units planned for 2026 (double 2025 output), delivering approximately 60% of H100 inference performance
• Ascend 920: 6nm process, exceeding 900 teraflops per card
• Ascend 950PR (debuted March 2026): 1.56 petaflops FP4, featuring Huawei's in-house HBM memory and claiming 2.8-times H20 performance
• Atlas 950 Supercomputer: 8,192 Ascend 950 processors, 16 exaflops performance
• Huawei's roadmap targets 4 zettaflops FP4 performance by 2028

Beijing's reciprocal talent controls are mirroring US chip export logic in reverse:

• State-secret travel restrictions originally placed on senior DeepSeek researchers have been extended to AI talent at Alibaba and other private firms, with Bloomberg first reporting the expansion in late May 2026
• Some AI professionals working on strategically important projects now require official government approval before traveling abroad
• The policy frames frontier-AI researchers themselves as restricted assets — a national-security treatment of human capital that parallels the US treatment of advanced chips
• Practical implications include constraints on international conference attendance, lab visits, and cross-border collaboration — a soft decoupling at the talent layer that compounds the already-decoupling at the supply layer

The semiconductor gap is real but narrowing. GLM-5's successful training entirely on Ascend hardware proves that Chinese labs can build frontier models without NVIDIA, and the DeepSeek efficiency story demonstrates that algorithmic innovation can partially compensate for hardware constraints. Beijing's reciprocal talent controls compound the same logic at the human-capital layer — what started as a one-way US export-control regime is hardening into a two-way decoupling that binds frontier-AI talent on both sides.

Key Takeaways

• Chinese foundation models are genuinely competitive — not derivative of US models, independently built under hardware constraints
• DeepSeek's current model is V4.1-Flash (September 10, 2026, MIT-licensed, 1 million-token context) — a 552 billion-parameter multimodal mixture-of-experts that activates 8 billion parameters per token while reading input. It replaced the April 2026 pair: V4-Flash retired the same day, and V4-Pro's endpoint routes to V4.1-Flash from September 14. V4-Pro is a 1.6 trillion total / 49 billion active mixture-of-experts model — no longer the largest open-weight release, having been passed by Kimi K3 and Qwen3.8 Max, but still the largest under a fully permissive license, and still downloadable even though DeepSeek has stopped serving it. Note the shape of the change: a smaller model displaced a larger one on cost and speed rather than on scale. DeepSeek closed its first outside venture round in June 2026, raising roughly 50 billion yuan (about $7.4 billion) from Tencent, CATL, NetEase, and JD.com, with founder Liang Wenfeng contributing 20 billion yuan himself and publicly committing the lab to open-source AGI as its core goal. R2 never shipped as a standalone model — that reasoning work went into V4
• DeepSeek V3.2 (MIT license, previous generation) was the model that first matched frontier performance and was trained for a fraction of comparable US model costs — its January 2025 release triggered a $589 billion single-day NVIDIA crash that permanently changed AI economics assumptions
• Data privacy is a critical distinction: Chinese-hosted APIs transmit data to Chinese servers subject to PRC law; running open-source weights locally eliminates this concern. Multiple countries have banned DeepSeek's hosted API on government devices
• The Qwen family (100+ languages, 1 million context, open weights through 3.6 plus proprietary Qwen3.7 and 3.8 Max flagships), Kimi K2.5 (1 trillion params), ERNIE 5.0 (2.4 trillion multimodal), and GLM-5 (744 billion, Ascend-only training) represent rapid advancement across the Chinese AI ecosystem
• China's Cyberspace Administration cleared Apple Intelligence on July 15, 2026 with Alibaba's Qwen as the model powering it for Chinese users and Baidu contributing a smaller model — the most consequential distribution win any Chinese model has achieved, and a case where regulatory geography rather than benchmark scores decided which model ships
• Moonshot AI shipped Kimi K3 on July 16, 2026 — a 2.8 trillion parameter mixture-of-experts model activating 16 of 896 experts per token, with a 1 million token context and native vision. It is the largest open-weights model any lab has published, and it beats Claude Opus 4.8 on most published coding benchmarks (FrontierSWE: 81.2 against 66.7). The weights went public on July 27, 2026 under a custom Kimi K3 License that is notably more restrictive than the K2 family's Modified MIT — commercial inference hosting above $20 million in revenue requires a separate Moonshot agreement
• Tencent's Hy4 preview (August 28, 2026) is the counterexample that makes the licensing rule concrete — 770 billion total parameters, roughly 49 billion active per token, context past one million tokens, published under plain Apache 2.0 with no territorial, revenue, or user limits. Every other lab on this page tightened terms on a newer or larger checkpoint; Tencent loosened them on its largest model, while its own HunyuanOCR, updated the same day, still ships under a Community License that excludes the European Union, the United Kingdom and South Korea outright. The license belongs to the checkpoint, not the vendor
• The open-weights concession is narrowing, not gone. Moonshot concedes K3 still trails Claude Fable 5 and GPT-5.6 Sol overall and names user experience as the gap; on HLE-Full reasoning it scores 43.5 against Fable 5's 53.3. And at $15 per million output tokens, K3 is priced like a frontier model, not as a discount play
• Xi Jinping made open source official national strategy at the World AI Conference in Shanghai on July 17, 2026 — the first time China's head of state framed open weights as industrial policy — one day after Kimi K3 shipped as the proof point. The strategic logic: if the most capable freely-available models are Chinese, the world's developers build on Chinese foundations, and weights cannot be stopped at a border the way chips can
• Moonshot AI previously shipped Kimi K2.6 + a $2 billion raise at a $20 billion valuation on May 7, 2026, with $200 million annualized recurring revenue. Combined with DeepSeek's own maiden round closing the following month, the two raises signalled China's open-weights labs were becoming the primary cost-pressure on US frontier API pricing
• ByteDance Doubao is China's most-used AI chatbot with 100 million+ DAU, while MiniMax and Zhipu AI have IPO'd in Hong Kong
• Huawei Ascend chips are advancing rapidly toward NVIDIA parity, with a roadmap to 4 zettaflops by 2028
• Beijing has extended state-secret travel restrictions originally placed on senior DeepSeek researchers to AI talent at Alibaba and other private firms (Bloomberg, late May 2026), requiring official approval before international travel — a reciprocal-controls signal that mirrors US chip export logic in reverse and hardens a two-way decoupling at the human-capital layer