Convika

หน้านี้ยังไม่ได้แปล จึงแสดงเป็นภาษาอังกฤษ

Chinese AI models in November 2026: the latest from DeepSeek, Qwen, Kimi, GLM, and MiniMax

Quick answer

As of November 20, 2026, the newest models from China's five best-known AI labs are DeepSeek-V4.1-Flash-1017, Alibaba's Qwen4-Max (on Qwen Cloud since November 16), Moonshot AI's Kimi K3.1 Pro, Z.ai's GLM-5.3, and MiniMax M3.1 Flash. All five labs publish open weights on Hugging Face (Qwen's newest are Qwen3.8-Max's), DeepSeek's under the MIT license and the others' under licenses that add conditions for some commercial uses, and their own APIs charge $0.15 to $2 per million input tokens. DeepSeek's privacy policy says it stores data in China, Kimi's and Z.ai's API policies say Singapore, MiniMax's says the US, and Qwen Cloud lets you pick the region.

Publish my HTML page free

Free plan: no credit card, 1 published page, 10,000 page views a month.

On this page

Five Chinese labs make most of the models developers outside China ask about: DeepSeek, Alibaba's Qwen team, Moonshot AI (Kimi), Z.ai (Zhipu AI's GLM models), and MiniMax. Each released its current model between August and November 2026. Each sells API access and also publishes model weights you can download.

Every fact below comes from the lab's own model card, pricing page, changelog, license file, or privacy policy, read on November 16, 2026 (Japan time), and Qwen's again on November 20. Benchmark numbers are each vendor's own, with the vendor named. Vendors run them with different setups, so the numbers don't make one leaderboard. Where a vendor's pages don't say something, this guide says so. Model names and prices change every few weeks: check the vendor's page before you commit.

Chinese AI models list: the newest from each lab

Chinese AI models, checked November 16, 2026; Qwen checked November 20. Prices are each vendor's list price on its own API, in US dollars per million tokens, input / output.

Lab and newest model (released)Open weights and licenseContextAPI price per 1M tokensWhere to use it outside China
DeepSeek: DeepSeek-V4.1-Flash-1017 (October 17, 2026), API name deepseek-flash. The larger DeepSeek-V4-Pro-0813 came out August 13Yes, MIT1M tokens, up to 384K outFlash: $0.30 / $1.20 at peak hours, half that off-peak. V4-Pro: $1.32 / $3.96 at peak, half off-peakDeepSeek API, Microsoft Foundry, Google Cloud Model Garden (V4.1-Flash), OpenRouter, Hugging Face
Qwen (Alibaba): Qwen4-Max (November 16, 2026), the first Qwen4 model. Qwen3.8-Max came out August 3, snapshot qwen3.8-max-0902Qwen4-Max: no. Qwen3.8-Max: yes, as Qwen3.8-2.4T-A95B, under the Qwen3.8-Max LicenseQwen4-Max: 1,000,000 tokens. Qwen3.8-Max: 1M on Qwen Cloud; its open weights 262,144 native, extensible to 1,010,000Qwen4-Max: $1.50 / $4.50 in Singapore, $1.413 / $4.24 in Beijing and Global. Qwen3.8-Max: $2 / $6. Qwen3.8-Flash: $0.15 / $0.47Qwen Cloud (Singapore, Frankfurt, Tokyo, Virginia, Hong Kong), OpenRouter, Hugging Face (Qwen3.8). Microsoft Foundry has Qwen3.8-27B
Kimi (Moonshot AI): Kimi K3.1 Pro (October 2026); weights on Hugging Face November 2Yes, under the Kimi K3.1-Pro License1,048,576 tokens$2 / $10; cached input $0.20Kimi API, Amazon Bedrock (K3.1 Pro Highspeed), Microsoft Foundry, OpenRouter, Hugging Face
GLM (Z.ai, Zhipu AI): GLM-5.3 (August 18, 2026). GLM-5.3-Flash came out August 26GLM-5.3: yes, under the GLM-5.3 License. GLM-5.3-Flash: MIT1M tokens, up to 128K out, text onlyGLM-5.3: $1.40 / $4.40. GLM-5.3-Flash: $0.15 / $0.50Z.ai API, Amazon Bedrock (GLM 5.3), Google Cloud Model Garden (GLM 5.2), OpenRouter, Hugging Face
MiniMax: MiniMax M3.1 Flash (October 19, 2026)Yes, under the MiniMax Community License1M tokens$0.15 / $0.60MiniMax API, OpenRouter, Hugging Face

DeepSeek's peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Its pricing page says "Off-peak rates are half of the peak rates", and weekends and Chinese public holidays are off-peak in full. Amazon Bedrock also lists older DeepSeek (V3.2), Qwen (Qwen3), and MiniMax (M2.5) models, and Microsoft Foundry also lists DeepSeek-V4-Pro.

Who makes what: Qwen Cloud is Alibaba Cloud's Qwen service for markets outside China, launched in Singapore. Qwen Cloud's own model list calls GLM-5.3 "Zhipu AI's most powerful model for programming capabilities to date"; Z.ai sells it on its own API.

Other Chinese labs with open weights

  • Xiaomi: MiMo-V2.6 Pro, MIT license, on Hugging Face since September 21, 2026.
  • Meituan: LongCat-2.5, MIT license, since November 4, 2026.
  • StepFun: Step-5 Preview, "StepFun's flagship model for agentic work", with 600B total and 27B active parameters and up to 1M context, under the StepFun Model License 1.0, since September 20, 2026.
  • Baidu: ERNIE-5.5-Flash-Base, under Baidu's own ERNIE 5.5 Flash Base Model License Agreement, since October 21, 2026.
  • ByteDance (Seed 2.1 Turbo) and Tencent (HY4) list API models on OpenRouter. This guide didn't read their own pages.

Open-source vs closed Chinese AI models

All five labs publish weights you can download and run yourself. Only some do it under a standard open-source license. The license file in each Hugging Face repository says what you may do with the weights:

WeightsLicenseConditions in the license
DeepSeek-V4.1-Flash-1017, DeepSeek-V4-Pro-0813MITKeep the copyright notice. The card: "This repository and the model weights are licensed under the MIT License."
GLM-5.3-FlashMITKeep the copyright notice
Qwen3.8-27BApache 2.0The Apache 2.0 terms
GLM-5.3GLM-5.3 LicenseA "Model as a Service" business with more than US$10 billion in revenue over 12 months "must pass Z.AI's security review" before commercial use
Kimi K3.1 ProKimi K3.1-Pro LicenseA product with more than 100 million monthly users or US$20 million in monthly revenue must show "Kimi K3.1-Pro" on its interface. A Model-as-a-Service business above US$20 million in revenue over 12 months needs a separate agreement. Neither applies to internal use
Qwen3.8-2.4T-A95B (Qwen3.8-Max)Qwen3.8-Max LicenseThe same display rule at 100 million monthly users or US$20 million monthly revenue. A "Model as a Service" or "AI Work Assistant" business above US$50 million in revenue over 12 months needs a separate license from Qwen
Qwen3.8-Flash-NextQwen Community License 1.0The same display rule. Any "Model as a Service" or "AI Work Assistant" business needs a separate license for commercial use, whatever its revenue
MiniMax M3.1 FlashMiniMax Community LicenseCommercial use must show "Built with MiniMax M3.1-Flash" and send MiniMax a one-time notice. Above US$20 million in yearly revenue, it needs written authorization first. An appendix bars uses including "any military purpose"

Qwen's newest models stay closed. Qwen4-Max runs on Qwen Cloud, through Token Plan or the pay-as-you-go API, and Qwen's Hugging Face page lists no Qwen4 weights on November 20, 2026. Qwen4-Max-Preview runs on Token Plan until 24:00 on November 25, 2026 (UTC+8). And the hosted Qwen3.8-Max does more than its open weights: Qwen's card calls it "the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools."

The Qwen 3.8 family: Max, 27B, and Flash Next

  • Qwen3.8-2.4T-A95B is the open release of Qwen3.8-Max: "2.4T in total and 95B activated" parameters. Qwen writes: "For the first time, Qwen3.8 brings a Qwen-Max-class model to open release."
  • Qwen3.8-27B is the small model under Apache 2.0, on Hugging Face since August 5, 2026.
  • Qwen3.8-Flash-Next is, in Qwen's words, an "experimental preview of the architecture that will underpin Qwen4": 125B parameters with 6B activated, plus 51B of n-gram embedding. The hosted Qwen3.8-Flash is "the official version based on Qwen3.8-Flash-Next".

Free ways to use Chinese AI models

Download the weights. Each repository above is free to download under its license. The parameter counts show the hardware it takes to run them. Kimi's card gives K3.1 Pro 2.8T total and 104B activated parameters. MiniMax gives M3.1 Flash "Approximately 428B total; 23B activated per token", and Z.ai gives GLM-5.3-Flash "320B total parameters and 18B activated". DeepSeek's model card doesn't give a parameter count.

Use a free API tier.

  • Z.ai's pricing page lists GLM-4.7-Flash and GLM-4.5-Flash as "Free" for input and output. They are older models, not GLM-5.3.
  • Qwen Cloud gives new users in its Singapore region a free quota, "typically 1,000,000 tokens" per model, valid for 90 days. It starts once you add payment information. "Other regions have no free quota."
  • OpenRouter lists no free variant of these five labs' models on November 16, 2026.

Use the chat app.

  • Kimi's pricing page lists a Free plan at $0, next to paid plans from Plus at $19 a month, or $180 a year.
  • chat.z.ai opened for this guide as a guest, without signing in, with GLM-5.3-Flash selected.
  • chat.qwen.ai opened with Log in and Sign up buttons and Qwen3.7-Plus selected. The page states no price.
  • DeepSeek's homepage links to its web chat and states no price there. chat.deepseek.com showed a bot check to this guide's browser.

Not free: Kimi's API says "Kimi K3.1 Pro is a flagship model: it is unlocked after a successful top-up (minimum $1)." DeepSeek's API deducts fees "from your topped-up balance or granted balance".

Chinese AI models for coding: DeepSeek vs Qwen vs Kimi

Each lab pitches its newest model at coding agents, and each reports a score on DeepSWE v1.1, a coding-agent benchmark. Each ran it with its own harness and settings. Read each number as that vendor's claim, not as a ranking. Qwen's row is Qwen3.8-Max, the model its card scores; Qwen4-Max came out on Qwen Cloud on November 16.

ModelWhat the maker saysThe maker's own DeepSWE v1.1 score, and its setupTools the maker names
DeepSeek-V4.1-Flash-1017"substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks", with shorter run times and outputDeepSeek reports 77.0, with the Mini-SWE-V2 harnessClaude Code, GitHub Copilot, OpenCode, and its own DeepSeek Harness. The API takes OpenAI and Anthropic formats
Qwen3.8-Max"substantial gains across coding, professional work, research, and long-horizon agentic tasks"Qwen reports 56.6, the higher of its Claude Code and mini-SWE-agent runsClaude Code and Codex, per the Qwen3.8-Flash page. Qwen Cloud has OpenAI- and Anthropic-compatible endpoints
Kimi K3.1 Pro"covers the full range of coding work, from quick fixes in real codebases to full-project builds"Moonshot reports 75.5, run in Claude CodeKimi Code. The API is "OpenAI/Anthropic-compatible"
GLM-5.3"the most capable open-weights model for coding, with a 50% improvement over GLM-5.2 on our in-house Z.ai Code Bench"Z.ai reports 66.9, with mini-swe-agentClaude Code, Kilo Code, Cline, and OpenCode, per Z.ai's docs
MiniMax M3.1 Flash"A generational leap in agentic tasks, coding, and cybersecurity"MiniMax reports 73.7, with mini-SWE-agent 2.4.5 at 512K contextMiniMax Code

How DeepSeek, Qwen, and Kimi differ for coding

  • Price per run. A long agent run spends mostly output tokens. DeepSeek's Flash costs $1.20 per million output tokens at peak and $0.60 off-peak. Qwen4-Max costs $4.50 in Singapore, Qwen3.8-Max $6, and Kimi K3.1 Pro $10.
  • Images. DeepSeek's pricing page marks vision as supported on Flash and "Not supported" on V4-Pro. Kimi K3.1 Pro takes text and images. Qwen3.8-Max on Qwen Cloud takes text, images, and video.
  • Running it yourself. DeepSeek's weights are MIT. Kimi's and Qwen's licenses add the display and Model-as-a-Service conditions in the table above.
  • Against US models. Moonshot's own table gives Kimi K3.1 Pro 65.7 on Terminal-Bench 4.0, Claude Opus 5.5 66.4, and GPT-6 Astra 58.2, each at max reasoning effort.

Each lab's coding benchmarks side by side, with Claude Code, Cursor, Cline, and Ollama setup: DeepSeek vs Qwen vs Kimi for coding.

Kimi's K3.1 Pro post shows front-end pages and Three.js scenes the model built end to end, and Qwen scores its models on building React projects with its in-house QwenReactBench. When one of these models gives you a page as a single HTML file, put it online with Convika's publish tool. Paste the code or drop the file, click Get my preview link, and you get a private link that works for 3 minutes, with scripts and forms off. Publish it from a free account to keep it at one address. In the dashboard, Add a form to my page adds a form that keeps the entries; Free shows 50 form entries a month.

Chinese vs US AI models: release dates, prices, and open weights

Current models, checked November 16, 2026; Qwen4-Max November 20. List prices per million tokens, input / output, on each lab's own API. "Weights published" means the lab has put that model's weights on Hugging Face. The US flagships themselves, with plans and benchmarks, are compared in the best AI models right now.

ModelLabReleasedAPI price per 1M tokensWeights published
DeepSeek-V4.1-Flash-1017DeepSeekOctober 17, 2026$0.30 / $1.20 at peak, half off-peakYes
Qwen4-MaxAlibaba (Qwen)November 16, 2026$1.50 / $4.50 in SingaporeNo
Qwen3.8-MaxAlibaba (Qwen)August 3, 2026$2 / $6Yes
Kimi K3.1 ProMoonshot AIOctober 2026$2 / $10Yes
GLM-5.3Z.aiAugust 18, 2026$1.40 / $4.40Yes
MiniMax M3.1 FlashMiniMaxOctober 19, 2026$0.15 / $0.60Yes
GPT-6 AstraOpenAISeptember 3, 2026$10 / $50No
GPT-6.1 SolOpenAISeptember 29, 2026$2 / $10No
Claude Fable 5.2AnthropicNovember 9, 2026$8 / $40No
Claude Opus 5.5AnthropicSeptember 22, 2026$4 / $20No
Gemini 4 CarbonGoogleNovember 10, 2026$2 / $10 through November 30, then $4 / $20No

OpenAI's prices are its standard rate for shorter prompts. Token counts differ between models, so one price per token doesn't mean one price per task. Anthropic's pricing page says its newer tokenizer "produces approximately 30% more tokens for the same text."

Release cadence. Every model in this table came out between August 3 and November 16, 2026. DeepSeek's changelog lists 6 model releases from April 24 to October 17, 2026. Z.ai shipped GLM-5 on February 12, GLM-5.1 on April 7, GLM-5.2 on June 16, and GLM-5.3 on August 18. Moonshot's changelog adds Kimi K2.6 in April, K2.7 Code in June, and K3 in July, before K3.1 Pro in October.

Prices. At list price, Kimi K3.1 Pro costs the same per token as GPT-6.1 Sol, and Gemini 4 Carbon matches both until November 30. The lowest output prices in the table are MiniMax M3.1 Flash ($0.60) and DeepSeek's Flash ($1.20 at peak). The highest are GPT-6 Astra ($50) and Claude Fable 5.2 ($40).

Open weights. Every Chinese model in this table but Qwen4-Max has weights on Hugging Face. None of the US models do. OpenAI's main open-weight models there, gpt-oss-120b and gpt-oss-20b, are from August 2025. Google's are Gemma 4, from 2026, and Meta's newest Llama 4 repositories are from April 2025. Anthropic's Hugging Face page lists no models.

Benchmarks against US models. Each Chinese lab's card compares its model with US models, but often with the previous generation. DeepSeek reports 77.0 on DeepSWE v1.1 for V4.1-Flash-1017, against 73.0 for GPT-5.6 Sol and 74.0 for Claude Opus 5. Qwen reports 67.7 on SWE-bench Pro for Qwen3.8-Max, against 69.2 for Claude Opus 4.8, 80.0 for Claude Fable 5, and 64.6 for GPT-5.6 Sol. Z.ai reports 88.2 on Terminal Bench 2.1 for GLM-5.3, against 88.8 for GPT-5.6 Sol. Moonshot's table is the one that includes GPT-6 Astra and Claude Opus 5.5.

Data and compliance: where each vendor processes your data

Each vendor's privacy policy or API terms says where your data goes. This section quotes them; it isn't legal advice.

APICompany in the termsWhere data is stored or processedAPI content and model training
DeepSeekHangzhou DeepSeek Artificial Intelligence Co., Ltd., "with its registered address in China""we directly collect, process and store your Personal Data in People's Republic of China"The policy lists "to train and improve our technology, such as our machine learning models and algorithms"
Qwen CloudIntelligent Cloud Computing (Singapore) Private Limited"Where request data is stored: The region whose endpoint you call." Singapore runs inference on "Nodes outside the Chinese mainland". In Frankfurt, Hong Kong, Tokyo, and Virginia, the default workspace's Global scope uses "Any available node, including the Chinese mainland and other regions"Qwen Cloud doesn't say: its customer agreement doesn't mention training
KimiMoonshot AI PTE. LTD., in Singapore"We store the information we collect in secure servers located in Singapore.""Unless otherwise expressly agreed in writing, Customer Content may be used" to improve the Services. Customers who want limits on training "may contact Moonshot AI to discuss available enterprise arrangements"
Z.aiJINGSHENG HENGXING TECHNOLOGY PTE.LTD, in Singapore"Customer Data is generally processed in Singapore.""The Company do not store any of the content the Customer or its End Users provide or generate". It is "processed in real-time" and "not saved on our servers"
MiniMaxNanonoble Pte. Ltd., in Singapore"Your personal data described in this Policy are stored in the data center located in the United States"MiniMax doesn't say: its API privacy policy doesn't cover training on API content

Three more points from the same documents:

  • Qwen Cloud's scope is set once. To keep inference in one place, create a workspace with the local scope: EU, Japan, US, or Hong Kong only. Qwen's docs: "A workspace's scope is set when the workspace is created and cannot be changed afterward."
  • DeepSeek's outputs are yours to reuse. Its Open Platform Terms say you may use inputs and outputs for "training other models (such as model distillation)". Its November 13, 2026 changelog asks apps that serve the public through its new built-in tools to turn on its safety check and "fulfill obligations such as content safety review, large model registration and AI labeling."
  • Google Cloud keeps prompts away from the model's maker. Google's Model Garden page says: "Customer prompts and model responses are not shared with third parties when using the Gemini Enterprise API, including open models." This guide read only the model lists of Amazon Bedrock and Microsoft Foundry, not their data terms. If you run the weights on your own servers, your prompts stay there; the license still applies.

Frequently asked questions

What are the main Chinese AI models?

As of November 20, 2026, the main Chinese AI models are DeepSeek-V4.1-Flash-1017 and DeepSeek-V4-Pro from DeepSeek, Qwen4-Max and Qwen3.8-Max from Alibaba, Kimi K3.1 Pro from Moonshot AI, GLM-5.3 from Z.ai (Zhipu AI), and MiniMax M3.1 Flash. Xiaomi (MiMo-V2.6), Meituan (LongCat-2.5), StepFun (Step-5 Preview), and Baidu (ERNIE 5.5) publish open-weight models too.

Which Chinese AI models are open source?

DeepSeek-V4.1-Flash-1017, DeepSeek-V4-Pro-0813, and GLM-5.3-Flash are under the MIT license, and Qwen3.8-27B is under Apache 2.0. Kimi K3.1 Pro, GLM-5.3, Qwen3.8-Max (as Qwen3.8-2.4T-A95B), and MiniMax M3.1 Flash publish open weights under their own licenses, which add conditions for some commercial uses. Qwen4-Max and Qwen4-Max-Preview have no public weights.

Are Chinese AI models free?

The weights are free to download, and some chat apps and API tiers are free. Kimi lists a Free plan at $0, chat.z.ai works as a guest, Z.ai's API offers GLM-4.7-Flash free, and Qwen Cloud gives new users in Singapore a free quota of "typically 1,000,000 tokens" per model for 90 days. Past those, the APIs charge per token.

Which Chinese AI model is best for coding?

No neutral ranking exists: each lab reports its own coding scores, run with its own setup, so the numbers don't line up. Each lab pitches its newest model at coding agents, and DeepSeek, Qwen, and Z.ai name Claude Code among the tools their models work in. Pick two or three by price and license, then test them on your own code.

DeepSeek vs Qwen vs Kimi: what's the difference?

They differ most in price and license. DeepSeek's Flash costs $0.30 / $1.20 per million tokens at peak, half off-peak, with MIT weights. Qwen4-Max costs $1.50 / $4.50 in Singapore with closed weights, and Kimi K3.1 Pro $2 / $10 under its own license. DeepSeek's privacy policy says it stores data in China, Kimi's API policy says Singapore, and Qwen Cloud lets you choose the region.

Are Chinese AI models cheaper than US models?

At list price, they cost less or the same. MiniMax M3.1 Flash costs $0.15 / $0.60 per million tokens and GLM-5.3 $1.40 / $4.40, against $4 / $20 for Claude Opus 5.5 and $10 / $50 for GPT-6 Astra. Kimi K3.1 Pro ($2 / $10), the priciest of the five, costs the same as GPT-6.1 Sol. Models count tokens differently, so compare the cost of a real task.

Where does DeepSeek store my data?

In China. DeepSeek's privacy policy says: "we directly collect, process and store your Personal Data in People's Republic of China." Its Open Platform Terms say the API is run by Hangzhou DeepSeek Artificial Intelligence Co., Ltd. To keep prompts away from DeepSeek, run its MIT weights on your own servers, or use DeepSeek-V4.1-Flash on Google Cloud, whose page says prompts and responses "are not shared with third parties".

What is the newest Qwen model?

Qwen4-Max, the first Qwen4 model, on Qwen Cloud since November 16, 2026, through Token Plan or the pay-as-you-go API. It costs $1.50 / $4.50 per million tokens in Singapore and $1.413 / $4.24 in Beijing and Global, with a 1,000,000-token context and no input-length tiers. Its weights aren't on Hugging Face. Qwen's newest open weights are Qwen3.8-2.4T-A95B (Qwen3.8-Max, August 3, 2026), with Qwen3.8-27B and Qwen3.8-Flash-Next.

What is the latest Chinese AI model?

Qwen4-Max, from Alibaba, on Qwen Cloud since November 16, 2026. Before it came MiniMax M3.1 Flash (October 19), DeepSeek-V4.1-Flash-1017 (October 17), and Kimi K3.1 Pro (October 2026, weights on Hugging Face November 2). Meituan's LongCat-2.5 open weights came out November 4.

Can I use Chinese AI models in the US?

Yes, in several ways. Qwen Cloud has a US (Virginia) region with a US-only scope, and OpenRouter lists all five labs' newest models. Amazon Bedrock lists GLM 5.3 and Kimi K3.1 Pro Highspeed, Microsoft Foundry lists DeepSeek-V4.1-Flash and Kimi-K3.1-Pro, and Google Cloud's Model Garden lists DeepSeek-V4.1-Flash and GLM 5.2. You can also run the open weights on your own servers.

Sources

Each page was read on November 16, 2026, between 08:29 and 09:00 Japan time, and Qwen's Qwen4-Max pages on November 20; the dates in brackets are the page's own.

For the older and retiring model ids on these clouds, see AI model retirements, November 2026 to February 2027.

Next step

How to password protect an HTML page or website

A JavaScript password script leaves the password in the page source. To password protect an HTML page or link, encrypt the file or let a server check it.

Related articles

More ways to send your page

Skipbox

Cheap & fast GitHub Actions runners, unlimited for $39 a month

Your own machine keeps Docker layers and package caches between jobs. Switch by changing one runs-on line.

Start a free Skipbox trial⁠

Share this article

Put the page your AI model built online

Paste the HTML that DeepSeek, Qwen, Kimi, or any other model wrote. Convika publishes it at a lasting link, keeps the form entries, and shows where visitors came from.

Publish my HTML page free

Free plan: no credit card, 1 published page, 10,000 page views a month.