Chinese AI models in 2026: the top language, video and image models, checked in November

This is a working reference to the Chinese AI models that matter in November 2026: who makes each one, the latest version I could confirm, whether the weights are open, how to use it from outside China and what it costs. Everything below was checked on 4 November 2026. Chinese labs ship new versions most months, so treat every version number here as a snapshot with that date on it.

It covers language models first, then video and image models, then coding, the three leaderboards people quote most, and what all of this means for search visibility and data protection. How assistants find and cite sources in general is covered in the AI search guide.

Xiaomi's MiMo-V2.6 models and Ant Group's Ling 3.1 Flash are covered at the end of the next section. Both rank well among open-weight models.

DeepSeek

DeepSeek's current family is V4. It was previewed on 24 April 2026 as two open models: V4-Pro, with 1.6 trillion total and 49 billion active parameters, and V4-Flash, with 284 billion total and 13 billion active, both with a context of one million tokens.²⁴ V4-Pro reached general availability as DeepSeek-V4-Pro-0813 on 13 August. V4.1-Flash followed on 10 September with native image understanding, and an updated build, DeepSeek-V4.1-Flash-1017, took over the deepseek-flash API name on 17 October.¹ The weights for both current models are on Hugging Face under the MIT licence.²²⁵

The API accepts both the OpenAI Chat Completions and the Anthropic formats, so it works with tools built for either. The older deepseek-chat and deepseek-reasoner names were scheduled for retirement on 24 July 2026, so any integration still calling them needs checking.¹ DeepSeek charges peak and off-peak rates, with off-peak at half price. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays.²⁶

Alibaba: Qwen

Alibaba launched Qwen3.8-Max on 3 August 2026 as its largest model so far: 2.4 trillion parameters with about 95 billion active, a context window of up to one million tokens, and API access for developers worldwide through Alibaba Cloud Model Studio.⁴ The weights followed as Qwen3.8-2.4T-A95B. Its model card says the hosted Qwen3.8-Max is based on these weights and adds features such as vision input and a one million token default context. The card names the licence qwen3.8-max, so read it before any commercial use.⁵

Model Studio's recommended list now runs qwen3.8-max for complex reasoning and coding, qwen3.7-plus as the balanced option and qwen3.8-flash as the fast, cheap one.²⁷ Qwen 4 has not been released. At its Apsara Conference on 22 September 2026, Alibaba said Qwen 4 was in training and set out a roadmap for Qwen 4.5 and Qwen 5 at 5 to 10 trillion parameters,²⁸ but no release date had been published by 4 November.

Moonshot AI: Kimi

Moonshot released Kimi K3 on 16 July 2026.⁶ It is an open-weight model with 2.8 trillion total and 104 billion active parameters and a context of 1,048,576 tokens, published under the Kimi K3 License.⁸ Its successor, Kimi K3.1 Pro, came out on 24 October.⁶ Moonshot calls it its most capable 3T-class open-weight model, offers it on Kimi.ai, the Kimi apps, Kimi Work, Kimi Code and the API, and said the full weights would be released by 4 November.⁷ I could not confirm they were public when this was written.

Caixin reported API prices for K3.1 Pro of 13 yuan per million uncached input tokens and 68 yuan per million output tokens.²⁹ Moonshot's pricing page says the older K2 series closed to new users on 31 October 2026 and is due to be retired at the end of December, so anyone still on K2 should plan a move.³⁰

Zhipu AI: GLM and Z.ai

Zhipu AI's GLM models are sold internationally through the Z.ai developer platform, priced in US dollars.¹¹ The GLM-5 family has moved quickly this year: GLM-5 on 12 February, GLM-5.1 on 7 April, GLM-5.2 on 16 June with very long context support, GLM-5.3 on 18 August with a focus on coding and cybersecurity, and GLM-5.3-Flash on 26 August.⁹ The GLM-5.3 and GLM-5.3-Flash weights are on Hugging Face.¹⁰

MiniMax

MiniMax released MiniMax M3 on 1 June 2026 for agentic reasoning, tool use and coding, and on 19 October added MiniMax M3.1 Flash, which it describes as an ultra-efficient open-weight model for everyday work.¹² The M3.1 Flash model card lists about 428 billion total and 23 billion active parameters, a context of up to one million tokens and a minimax-community licence.¹³ M3 costs $0.30 per million input tokens and $1.20 per million output tokens for prompts up to 512,000 tokens, and M3.1 Flash costs $0.15 and $0.60.¹⁴ MiniMax also makes the Hailuo video models, covered below.

ByteDance: Doubao and Seed

ByteDance's Seed team builds the models behind the Doubao assistant. Seed2.1, released on 23 June 2026, is pitched as an agent model for real-world productivity with stronger end-to-end coding, and Doubao and Volcano Engine users can use it.¹⁵ The announcement says nothing about open weights. The Seed team's blog is also where it publishes its research, from the Seed 2.0 launch in February 2026 to work such as Seed Prover 1.5 and Seed Diffusion.³¹

Outside China, ByteDance's consumer assistant is Dola, which began as Cici in August 2023 and passed 10 million daily active users at the end of 2025.¹⁶ ByteDance's model pages link to Dola for consumers and to a BytePlus playground for developers.³² I could not find a published international price for Seed2.1.

Tencent: Hy, formerly Hunyuan

Tencent's Hunyuan models now go by the name Hy. The final version of Hy3 launched on 6 July 2026, after a preview in April.³³ It is a 295 billion parameter mixture-of-experts model with 21 billion active parameters and a 256K context, released under Apache 2.0.¹⁸ Hy4 launched on 4 November 2026 as a 770 billion parameter model that adds image understanding and computer use. It is available in Tencent's own products and through third-party platforms including OpenRouter. Standard API pricing is $0.834 per million input tokens and $2.501 per million output tokens, and Tencent says the weights and a system card will follow shortly.¹⁷

Baidu: ERNIE

Baidu released ERNIE 5.1 on 9 May 2026. People can chat with it on the ERNIE website and try it in Baidu AI Studio, and Baidu says it placed fourth globally and first among Chinese models on Arena's Search leaderboard, with a score of 1,223.¹⁹ I found no ERNIE 5 weights on Baidu's Hugging Face page, where its recent open releases are ERNIE 4.5 models and the ERNIE-Image family,²⁰ and ERNIE 4.5 is under Apache 2.0.²¹ I could not find an official international API price for ERNIE 5.1.

StepFun

StepFun's newest model is Step 5 Preview, a sparse mixture-of-experts model with 600 billion total and 27 billion active parameters and up to one million tokens of context. The weights are on Hugging Face under the StepFun Model License 1.0, which adds attribution and commercial-licence conditions to Apache-style terms, and the API runs on StepFun's own platform.²² OpenRouter lists it at $1 per million input tokens and $2.70 per million output tokens.²³

Xiaomi, Ant Group and Meituan

Three companies outside the usual list of Chinese AI labs now have models on the leaderboards. Artificial Analysis names Xiaomi's MiMo-V2.6-Pro and Ling 3.1 Flash 1019 as the next-best open models after DeepSeek and GLM.³⁴ Ling comes from InclusionAI, which describes itself as the home of Ant Group's artificial general intelligence work.³⁵ Meituan's LongCat-2.0 sits on SuperCLUE's Chinese leaderboard, ranked level with ERNIE 5.1.³⁶

The price per token only tells part of the story, because some models use far more tokens than others to finish the same job. Artificial Analysis publishes a cost per task for each model across its evaluations: DeepSeek V4.1 Flash 1017 (max) came to $0.60, Kimi K3.1 Pro to $3.68, GLM-5.3 (max) to $5.01 and Qwen3.8 Max (0902) to $10.02.³⁸ For a team running agents all day, that column matters more than the list price.

Video is where Chinese labs are strongest. On Artificial Analysis's text-to-video leaderboard, AA-Video-T2V v2.1, which ranks video with audio by people's votes between pairs of clips, Alibaba's Wan 3.0 was first, MiniMax H3 (768p) third and Dreamina Seedance 2.5 fifth when I checked. Kling 4.0 had not yet appeared there, and Kling 3.0 Omni 1080p (Pro) ranked 18th.³⁹

MiniMax's H3 model card links to the Hailuo AI app as one place to use it, and describes clips of 4 to 15 seconds at 24 frames a second with native stereo audio, plus a separate stage that regenerates output at 2K.⁴² The older Hailuo 2.3 and Hailuo-02 models now sit in a legacy section of MiniMax's price list.¹⁴ Shengshu, the company behind Vidu, says it grew out of Tsinghua University's AI research labs.⁴⁷

Chinese AI video releases in 2026

• February: Kuaishou launched the Kling 3.0 series, with clips of up to 15 seconds and native audio, first to Ultra subscribers.⁵⁰
• 31 July: MiniMax released H3, and ByteDance announced Seedance 2.5.¹²⁴³
• 24 August: Alibaba launched Wan 3.0 with closed weights, a break from its earlier practice of open-sourcing Wan models.⁴⁵
• 7 October: Shengshu released Vidu Q4 Preview.⁴⁷
• 15 October: Kling 4.0 launched.⁴⁰

The newest open video model from Tencent that I could confirm is still HunyuanVideo 1.5, from November 2025.⁴⁸ Whether a business should pay for any of these yet is covered in AI video generation is a waste of money right now.

In image generation, Chinese models trail the leaders by a wider margin than in video. On Artificial Analysis's text-to-image leaderboard, AA-Image-T2I v2.1, the first Chinese entries were ByteDance's Seedream 5.0 Pro in 16th place and Alibaba's Qwen-Image-3.0-Pro in 17th, behind OpenAI's GPT Image models and Google's Nano Banana models, among others.⁵¹

The release months for the two Qwen-Image rows come from Artificial Analysis's listing.⁵¹ Tencent's HunyuanImage 3.0 has 80 billion total and 13 billion active parameters and handles both text-to-image and image-to-image work.⁵⁵ Qwen-Image-2.1 has a 7 billion parameter image generation component, and can also produce transparent images and edit with up to 10 reference images.⁵³

On the public evidence, Moonshot's Kimi models are the strongest Chinese option for coding right now, with Qwen3.8-Max, DeepSeek V4.1 Flash and GLM-5.3 close behind. None of them leads the overall coding rankings, which are still headed by Anthropic, OpenAI and Google models, as the figures below show.

• Arena's WebDev leaderboard, which ranks models on front-end web development from user votes: Kimi K3.1 Pro (Max) was ninth with 1,695, marked preliminary. Qwen3.8-Max (0902) was 13th, DeepSeek-V4.1-Flash-20261017 (Max) 16th, Kimi K3 (Max) 18th, Tencent's Hy4 preview 23rd and GLM-5.3 (Max) 27th. Claude Opus 5.5 (Max) led with 1,810.⁵⁷
• SuperCLUE's agentic coding score in its July 2026 round: GLM-5.3 76.84, Kimi K3 75.79, DeepSeek-V4.1-Flash 71.58, Qwen3.8-Max-0902 70.53 and Doubao-Seed-2.1-pro 51.58.³⁶
• Terminal-Bench 2.1, as reported by each lab: Kimi K3 88.3,⁸ GLM-5.3 88.2,¹⁰ DeepSeek-V4-Pro-0813 87.9,²⁵ Qwen3.8-Max 86.6⁵ and Tencent's Hy3 71.7.¹⁸ Each lab ran its own tests, so treat these as claims.
• Kimi K3.1 Pro on Terminal-Bench 4.0: Moonshot reports 65.7%, narrowly behind Claude Opus 5.5 at 66.4%.⁷

On those numbers I would pick Kimi K3.1 Pro for the best Chinese model, if you are happy to use Moonshot's API, which also runs inside its own Kimi Code tool.⁷ I would pick DeepSeek V4.1 Flash where cost matters most, and GLM-5.3 for strong open weights to run yourself. DeepSeek is also the easiest to try in an existing setup, because its API accepts the Anthropic message format as well as OpenAI's,¹ and its October release notes name Claude Code and Codex among the tools that can call its server-side search.³

The three leaderboards people quote most measure different things, which is why they rarely agree on which Chinese model is best.

SuperCLUE

SuperCLUE is the large-model successor to CLUE, a Chinese language benchmark launched in 2019. In its July 2026 general leaderboard, models within one point of each other are treated as tied, and international models are left out of the ranking.³⁶ That makes it a good answer to which Chinese model is best in Chinese, and no help at all on whether a Chinese model beats GPT or Claude. Further down, Doubao-Seed-2.1-pro scored 65.94, Tencent's Hy3 62.13, MiniMax M3 56.90 and ERNIE 5.1 54.56. Its hallucination control scores ran from 63.17 to 89.73, with DeepSeek-V4-Pro-0813 the highest.³⁶

Arena, formerly LMArena

LMArena's leaderboards now live at arena.ai. Its text board ranks models from user votes. When I checked, Kimi K3 (Max) was the highest Chinese model in 19th place, followed by Qwen3.8-Max in 24th, GLM-5.3 (Max) in 28th and Xiaomi's MiMo-V2.6-Pro in 29th.⁵⁸ Kimi K3.1 Pro had reached the WebDev board but did not appear in the top 40 of the text board.

Artificial Analysis

Artificial Analysis runs its own tests. Its Intelligence Index v5.0.1 combines nine evaluations, including Terminal-Bench 4.0 and FrontierCode 2.0.³⁴ Kimi K3.1 Pro was the top Chinese model at 46, against 60 for the leader, and the next Chinese entries were Qwen3.8 Max (0902) at 34, DeepSeek V4.1 Flash 1017 (max) at 33 and GLM-5.3 (max) at 32.³⁸ Its open-weights page is where Chinese labs dominate: the six highest-scoring models in its default view came from DeepSeek, Z AI, Xiaomi, InclusionAI, MiniMax and StepFun, ahead of Mistral Large 4 and Meta's Muse Spark Open.³⁴

Figures quoted in the press can differ from the live page. Caixin reported a score of 55 for Kimi K3.1 Pro on 26 October,²⁹ while the page I read on 4 November showed 46 on version 5.0.1. Quote the live page and its index version, not a number from a launch article.

Open weights mean you can download a model and run it yourself. The licence decides what you may do with it, and some Chinese licences carry conditions that matter in the UK and EU.

• MIT: DeepSeek-V4-Pro-0813 and DeepSeek-V4.1-Flash-1017.²⁵²
• Apache 2.0: Tencent's Hy3¹⁸ and Baidu's ERNIE 4.5.²¹
• Company licences: the Kimi K3 License,⁸ minimax-community for MiniMax M3.1 Flash,¹³ the StepFun Model License 1.0 with its attribution and commercial-licence conditions,²² and qwen3.8-max for the Qwen3.8 weights.⁵ Read each one before building on it.
• A research licence: Qwen-Image-2.1 is released under the Qwen Research License Agreement, so check whether your use counts.⁵³
• Territory exclusions: the Tencent Hunyuan Community License, used by HunyuanVideo 1.5 and HunyuanImage 3.0, defines its territory as the world excluding the European Union, the United Kingdom and South Korea.⁴⁹⁵⁵ MiniMax's H3 model card links a separate application form for users in the USA, EU, UK and South Korea.⁴²

For a UK business, the territory point is the one that bites. As I read it, a licence whose territory excludes the UK gives you no right to use the model here, however easy it is to download. Ask a lawyer before relying on one of these models commercially.

If you sell to people in mainland China, Chinese assistants are where they ask. QuestMobile's first-half 2026 report counted 499 million monthly active users of AI-native apps in China by May 2026, with Doubao on 382 million, Qwen on 167 million and DeepSeek on 130 million.⁵⁹

Outside China the reach is smaller but real. ByteDance's Dola passed 10 million daily active users at the end of 2025, and third-party data cited by Pandaily put it among the top 20 free Google Play downloads in Indonesia, Malaysia, Mexico and the UK from October 2025.¹⁶ Chinese models also reach Western users without any Chinese app. Amazon Bedrock lists models from DeepSeek, Moonshot, MiniMax, Qwen and Z.ai,⁶⁰ and Hy4 was on OpenRouter at launch.¹⁷ A tool built on one of these models answers from its training data and whatever search it is connected to.

Search is where this gets specific. DeepSeek now offers what it calls an in-house search API, with Search and Fetch endpoints,¹ generally available since 17 October at $0.75 per 1,000 calls.³ I could not find a DeepSeek page documenting the crawler behind it. ByteDance's Bytespider is similarly undocumented, as I cover in the other AI crawlers in your logs. Without documentation there is no reliable way to allow or block these crawlers by name, so your server logs are the best evidence of what they fetch.

If China is a market for you, I would test your brand and category prompts in Doubao, Qwen and DeepSeek the same way you would test ChatGPT, using the method in how to measure visibility in AI search. If it is not, the work is the same as for any assistant: pages crawlers can read and consistent facts about your business across the web.

The data question depends on how you use a model. Using a Chinese company's own app or API sends your prompts to that company. Running open weights on your own servers, or through a cloud provider in your own region, does not involve the model's maker at all.

DeepSeek is the case regulators have acted on. Its privacy policy, last updated on 10 February 2026, says it collects, processes and stores personal data in the People's Republic of China, lists training its models among the uses of what people type, and offers an opt-out by email.⁶¹ These are the three regulatory actions I could confirm.

• Italy: on 30 January 2025 the Garante ordered an immediate limitation on DeepSeek's processing of Italian users' data and opened an investigation, after the two DeepSeek companies told it they did not operate in Italy and that European law did not apply to them.⁶²
• Germany: in June 2025 Berlin's data protection commissioner, Meike Kamp, reported the DeepSeek app to Apple and Google, saying it transferred users' data to China illegally and that DeepSeek had not provided convincing evidence the data was protected as EU law requires.⁶³
• South Korea: in April 2025 the Personal Information Protection Commission found that DeepSeek had sent device information and the text of users' prompts to Volcano, ByteDance's cloud platform. It recommended that DeepSeek establish a legal basis for overseas transfers and destroy the prompts it had sent. DeepSeek had paused new downloads in Korea since 15 February, which the commission described as a voluntary step.⁶⁴

All three concern DeepSeek's own service, and all date from 2025. None of them is about the open weights, which anyone can run without sending data to DeepSeek. These are the checks I would make before a team uses a Chinese model.