Chinese AI Landscape

The Chinese AI Model Guide: DeepSeek, Qwen, Ernie, and Beyond

Hong Kong AI Podcast/2026-03-07/12 min read/DeepSeekQwenERNIEZhipuMiniMaxStepFunChinese AI

Something remarkable happened in Chinese AI over the past two years. While the Western tech press focused on OpenAI and Anthropic, China's AI labs shipped dozens of frontier models — many of them open source, many of them genuinely excellent, and almost all of them accessible from Hong Kong without restrictions.

This guide covers the major Chinese AI models and companies as of early 2026. If you're a developer, researcher, or founder in Hong Kong, these aren't alternatives — for many of you, these are your primary tools.

DeepSeek

Company: DeepSeek (Hangzhou-based, founded 2023) Key models: DeepSeek-V3.2, DeepSeek-R1 License: MIT (fully open source) Access: API at api.deepseek.com + self-host via Hugging Face

DeepSeek made a splash with its reasoning and chat capabilities. In just over two years, they've gone from unknown to a household name in AI.

DeepSeek-V3.2 is their current flagship — a 671B parameter Mixture of Experts model with 37B active parameters per token. It won gold medals at both the 2025 IMO and IOI. MIT licensed, meaning you can use it for anything, including commercial applications.

DeepSeek-R1 is their reasoning model, comparable to OpenAI's o1. It excels at math and multi-step problem solving.

What makes DeepSeek special is the openness, not necessarily the raw quality. Everything is MIT licensed — you can download the weights, fine-tune them, deploy them commercially, and inspect the architecture. For coding specifically, DeepSeek can be slow and isn't the top performer — models like Claude Opus 4.6, MiniMax M2.5, and GLM-5 score higher on SWE-bench. But DeepSeek's price and self-hosting story are hard to beat.

Best for: Reasoning/chat tasks, self-hosting, cost-sensitive deployments, any application where MIT licensing and full model control matter.

Alibaba Qwen

Company: Alibaba Cloud (Qwen team) Key models: Qwen 3.5, Qwen 3-Max, Qwen 2.5-VL License: Apache 2.0 Access: Alibaba Cloud DashScope API, Hugging Face, chat.qwen.ai

Qwen is Alibaba's model family, and it's massive — over 100 open-weight models released to date. Qwen 3.5 is the current flagship: up to 397B parameters (MoE), supporting 201 languages.

The multilingual capability is Qwen's superpower. For Hong Kong developers building products that need both English and Chinese, Qwen is often the best choice. The model understands code-switching, Cantonese contexts, and the nuances of traditional vs. simplified Chinese.

Qwen 3.5 comes in multiple sizes, from tiny models you can run on a phone to the full 397B MoE. All Apache 2.0 licensed.

Hong Kong apps like 8BitOracle and SixLines are already running Qwen3 in production for bilingual chat as of March 2026. Local AI community members are also observing that the smaller Qwen 3.5 models are overperforming relative to their size — worth testing before reaching for the larger variants.

Best for: Bilingual applications, Chinese-language tasks, multilingual products, on-device deployment (smaller models).

Baidu ERNIE

Company: Baidu Key models: ERNIE 4.5, ERNIE 5.0 (preview) License: Apache 2.0 (4.5 series), closed (5.0) Access: Qianfan platform, Hugging Face

ERNIE was one of the first serious Chinese LLMs, and Baidu has continued pushing it forward. ERNIE 4.5 is a family of 10 model variants, from 0.3B dense to 424B MoE, all multimodal, all open-sourced under Apache 2.0.

ERNIE 5.0, announced in November 2025, is 2.4 trillion parameters and omni-modal (text, image, audio, video). As of early 2026, it's in preview and not yet open source.

Best for: Multimodal applications, Chinese search/knowledge tasks, teams already on Baidu's cloud platform.

ByteDance (Doubao)

Company: ByteDance Key models: Doubao-Seed-2.0 License: Partially open Access: Doubao app, Volcano Engine cloud

ByteDance's AI play is Doubao — both an app (200M+ users) and a model family. Doubao-Seed-2.0 is a family of four models, with the Pro variant scoring 98.3 on AIME 2025 and hitting a 3020 Codeforces rating.

Doubao is less developer-focused than DeepSeek or Qwen — it's more consumer-oriented, similar to how ChatGPT is the consumer face of OpenAI. But the underlying models are serious, and API access is available through Volcano Engine.

Best for: Consumer-facing applications, teams building on ByteDance's ecosystem.

Zhipu AI (Z.ai)

Company: Zhipu AI, rebranded internationally as Z.ai (July 2025) Key models: GLM-5, GLM-4.7, GLM-4.5 License: Open source (GLM-4.5) Access: API, IPO'd on HKEX January 8, 2026

Zhipu was co-founded by researchers from Tsinghua University and has positioned itself as China's "open-source champion." GLM-4.5 is a 355B MoE model described as the most advanced open-source MoE in China at its release.

GLM-5 is their latest — focused on autonomous coding and long-context tasks. Local practitioners in Hong Kong observed GLM-5 as being very strong for coding when it first appeared as "Pony Alpha" on OpenRouter, tested using OpenCode before the model's identity was even public. Zhipu's international rebrand to Z.ai and their HKEX IPO signal serious ambitions outside mainland China.

Best for: Research applications, coding tasks, long-context tasks, teams that value academic rigor.

Moonshot AI (Kimi)

Company: Moonshot AI (became China's fastest decacorn in 2026) Key models: Kimi K2.5, Kimi k1.5 License: Some models open-weight Access: API, kimi.ai

Kimi K2.5 might be the most interesting Chinese model you haven't heard of. It supports 262K context, has tool-use capabilities, and features an "Agent Swarm" system that spins up 100 parallel sub-agents using a technique called PARL.

Moonshot became China's fastest decacorn — their 20-day revenue after K2.5 launched exceeded their entire 2025 total. That's the kind of growth that gets attention.

Best for: Agent-based applications, tool-use workflows, long-context tasks.

MiniMax

Company: MiniMax (IPO'd on HKEX January 9, 2026) Key models: MiniMax-M2.5, MiniMax-M1 License: Modified MIT Access: API, Hugging Face

MiniMax-M2.5 is the model that made people do a double take: it scores 80.2% on SWE-Bench and matches Claude Opus on several benchmarks — at approximately 1/20th the cost. Released under a modified MIT license on Hugging Face.

The cost efficiency is the headline, but the model quality is the story. MiniMax went from relatively unknown to HKEX-listed in under a year.

Best for: Cost-sensitive production deployments, coding tasks, teams that need Claude-level quality without Claude-level pricing.

StepFun

Company: StepFun (seeking HKEX IPO) Key models: Step 3.5 Flash, Step 3 License: Apache 2.0 Access: API, Hugging Face, NVIDIA NIM

StepFun 3.5 Flash is an efficiency story. It's a 196B MoE model with only 11B active parameters per token — meaning it runs fast and cheap while scoring 97.3% on AIME 2025 and 74.4% on SWE-bench. All under Apache 2.0.

This is the model that makes the MoE architecture click: you get big-model quality with small-model speed. StepFun is reportedly seeking an HKEX IPO, continuing the trend of Chinese AI companies listing in Hong Kong.

Best for: Latency-sensitive applications, cost-efficient deployment, math and coding tasks.

SenseTime (SenseNova)

Company: SenseTime (headquartered in Hong Kong, listed on HKEX) Key models: SenseNova V6, SenseNova-MARS, SenseNova-SI License: Partially open source Access: API, some models on Hugging Face

SenseTime is Hong Kong's own AI giant — headquartered here and listed on HKEX with a ~$7.7B market cap. SenseNova V6 is their flagship (620B hybrid expert architecture). SenseNova-MARS (multimodal autonomous reasoning, open source) and SenseNova-SI (spatial intelligence, open source) target specific application domains.

Best for: Computer vision, spatial intelligence, teams that want to work with an HK-headquartered company.

How to Choose

The right model depends on your setup and what you're building:

-Using Cursor? Auto mode handles model routing. Some developers prefer Sonnet for quality but watch token usage.
-Using OpenCode? As of March 2026, MiniMax M2.5 is available free.
-Building bilingual products? Qwen 3.5 has strong multilingual support across 201 languages. See 8BitOracle and SixLines for HK-built examples.
-Optimizing for speed and cost? StepFun 3.5 Flash — frontier accuracy at a fraction of the cost.
-Building agent systems? Kimi K2.5 with its PARL agent swarm architecture.
-Want an HK-based provider? SenseTime SenseNova.
-On OpenRouter? Models rotate — check what's available. Pony Alpha (GLM-5) was previously a strong option.

The models are competitive. The licensing is permissive. And none of them will block you because of where you live.

Sources

This guide is updated regularly as new models release. Subscribe to the Hong Kong AI Podcast for the latest on AI tools and models available in Hong Kong.

Stay in the loop

Get notified when we publish new articles and episodes. No spam, just signal.

Something out of date or wrong? AI moves fast and we want to get it right. Let us know at contact@hongkongaipodcast.com