China

AI models, vendors, and hardware buying realities from China. Curated by myaihardware.com.

Models

Vendors

Languages

Why China matters for AI builders

China is the world's largest single market for AI deployment, meaning any hardware architecture not optimized for Chinese data-center power constraints and local GPU availability will struggle to achieve scale. Dominant model families like Qwen and DeepSeek define the current open-weight frontier, and their training and inference demands directly shape GPU procurement strategies from Huawei to NVIDIA. The underrated angle is that China's tight export controls on advanced logic chips force a unique reliance on edge computing and heterogeneous compute clusters, creating a hardware ecosystem that ultimately influences global supply-chain resilience.

Market context

US export controls limit Chinese access to advanced NVIDIA GPUs, while domestic alternatives like Huawei Ascend constrain raw compute. Labs like DeepSeek (V3/R1) and Alibaba (Qwen 2.5/3) release open-weight models frequently, often rivaling or surpassing top US closed models. Chinese firms compete by rapidly iterating on architecture and efficiency to establish dominance in the open-source LLM landscape.

Models (47 of 47)

Filter by vendor, license, or parameter range.

LicenseVendorParameter range

Qwen2.5-72B

Alibaba Cloud · China

Open

Params: 72B
Context: 131,072
Min VRAM: 48 GB
GGUF: Yes

Strong multilingual reasoning; High context length; SOTA on Chinese benchmarks

Qwen2.5-32B

Alibaba Cloud · China

Open

Params: 32B
Context: 32,768
Min VRAM: 24 GB
GGUF: Yes

Good balance of performance and size; Supports tool use

Qwen2.5-14B

Alibaba Cloud · China

Open

Params: 14B
Context: 32,768
Min VRAM: 12 GB
GGUF: Yes

Efficient for deployment; Good reasoning

Qwen2.5-7B

Alibaba Cloud · China

Open

Params: 7B
Context: 32,768
Min VRAM: 6 GB
GGUF: Yes

Lightweight, fast inference; Good for local deployment

Qwen2.5-Coder-7B

Alibaba Cloud · China

Open

Params: 7B
Context: 32,768
Min VRAM: 6 GB
GGUF: Yes

Specialized for code generation; Supports multiple programming languages

Qwen2.5-Coder-31B

Alibaba Cloud · China

Open

Params: 31B
Context: 32,768
Min VRAM: 24 GB
GGUF: Yes

Top-tier code generation; Large context window

Qwen2-VL-7B

Alibaba Cloud · China

Open

Params: 7B
Context: 32,768
Min VRAM: 8 GB
GGUF: No

Multimodal understanding (image+text); Strong OCR

Qwen2-VL-72B

Alibaba Cloud · China

Open

Params: 72B
Context: 131,072
Min VRAM: 60 GB
GGUF: Yes

High-fidelity image understanding; Long context with images

Qwen1.5-72B

Alibaba Cloud · China

Open

Params: 72B
Context: 32,768
Min VRAM: 48 GB
GGUF: Yes

Strong general knowledge; Efficient training

Qwen3-235B-A22B

Alibaba Cloud · China

Open

Params: 235B
Context: 131,072
Min VRAM: 48 GB
GGUF: Yes

Sparse activation for efficiency; Multilingual SOTA

Qwen3-30B-A3B

Alibaba Cloud · China

Open

Params: 30B
Context: 131,072
Min VRAM: 12 GB
GGUF: Yes

Very low active params per token; Fast inference on consumer GPUs

DeepSeek-V3

DeepSeek (High-Flyer) · China

Restricted

Params: 671B
Context: 131,072
Min VRAM: 300 GB
GGUF: Yes

Extreme scale, competitive with GPT-4; Efficient MoE with sparse activation

DeepSeek-R1

DeepSeek (High-Flyer) · China

Restricted

Params: 671B
Context: 131,072
Min VRAM: 300 GB
GGUF: Yes

Advanced reasoning with chain-of-thought; Reinforcement learning from human feedback

DeepSeek-V2

DeepSeek (High-Flyer) · China

Restricted

Params: 236B
Context: 131,072
Min VRAM: 160 GB
GGUF: Yes

Cost-effective MoE; Strong performance on math and code

DeepSeek-R1-Distill-Qwen-32B

DeepSeek (High-Flyer) · China

Open

Params: 32B
Context: 131,072
Min VRAM: 24 GB
GGUF: Yes

Distilled from R1, strong reasoning; Efficient for local deployment

DeepSeek-Coder-V2

DeepSeek (High-Flyer) · China

Restricted

Params: 236B
Context: 131,072
Min VRAM: 160 GB
GGUF: Yes

Top-tier code generation and completion; Support for 300+ languages

DeepSeek-Math-7B

DeepSeek (High-Flyer) · China

Open

Params: 7B
Context: 8,192
Min VRAM: 6 GB
GGUF: Yes

Specialized for mathematical reasoning; Strong on GSM8K and MATH

DeepSeek-VL-7B

DeepSeek (High-Flyer) · China

Open

Params: 7B
Context: 4,096
Min VRAM: 8 GB
GGUF: No

Multimodal (image+text); Strong on OCR tasks

Baichuan3-7B

Baichuan Intelligent Technology · China

Restricted

Params: 7B
Context: 8,192
Min VRAM: 6 GB
GGUF: Yes

Good Chinese language capabilities; Lightweight

Baichuan3-13B

Baichuan Intelligent Technology · China

Restricted

Params: 13B
Context: 8,192
Min VRAM: 10 GB
GGUF: Yes

Strong Chinese NER and classification; Balanced size

Baichuan4-7B

Baichuan Intelligent Technology · China

Restricted

Params: 7B
Context: 32,768
Min VRAM: 6 GB
GGUF: Yes

Updated with longer context; Efficient for fine-tuning

Baichuan-M1-14B

Baichuan Intelligent Technology · China

Restricted

Params: 14B
Context: 65,536
Min VRAM: 12 GB
GGUF: Yes

MoE for higher efficiency; Good balance of size and performance

ChatGLM3-6B

Zhipu AI · China

Open

Params: 6B
Context: 8,192
Min VRAM: 6 GB
GGUF: Yes

Excellent Chinese dialogue; Low resource requirement

GLM-4-9B

Zhipu AI · China

Open

Params: 9B
Context: 131,072
Min VRAM: 8 GB
GGUF: Yes

Very long context window; Strong Chinese and code tasks

GLM-4-32B

Zhipu AI · China

Open

Params: 32B
Context: 131,072
Min VRAM: 24 GB
GGUF: Yes

Excellent long-context handling; Strong overall performance

Yi-1.5-34B

01.AI · China

Open

Params: 34B
Context: 32,768
Min VRAM: 28 GB
GGUF: Yes

Competitive code and math; Good for fine-tuning

Yi-1.5-6B

01.AI · China

Open

Params: 6B
Context: 32,768
Min VRAM: 6 GB
GGUF: Yes

Fast inference; Good for edge deployment

Yi-Coder-9B

01.AI · China

Open

Params: 9B
Context: 65,536
Min VRAM: 8 GB
GGUF: Yes

Specialized for code; Long context window

InternLM2-20B

Shanghai AI Laboratory · China

Open

Params: 20B
Context: 32,768
Min VRAM: 16 GB
GGUF: Yes

Strong reasoning and math; Open-source friendly

InternLM2-7B

Shanghai AI Laboratory · China

Open

Params: 7B
Context: 32,768
Min VRAM: 6 GB
GGUF: Yes

Efficient for fine-tuning; Good Chinese NLP

InternLM2.5-7B

Shanghai AI Laboratory · China

Open

Params: 7B
Context: 65,536
Min VRAM: 6 GB
GGUF: Yes

Longer context than 2.0; Improved instruction following

MiniCPM-2B

OpenBMB (Tsinghua) · China

Open

Params: 2B
Context: 4,096
Min VRAM: 2 GB
GGUF: Yes

Ultra-compact, fast on CPU; Good for mobile devices

MiniCPM-3-4B

OpenBMB (Tsinghua) · China

Open

Params: 4B
Context: 8,192
Min VRAM: 4 GB
GGUF: Yes

Better performance than 2B; Still lightweight

MiniCPM-V-2.6

OpenBMB (Tsinghua) · China

Open

Params: 8B
Context: 8,192
Min VRAM: 8 GB
GGUF: No

Multimodal in a small package; Good for edge vision tasks

Hunyuan-Large

Tencent · China

Open

Params: 389B
Context: 131,072
Min VRAM: 200 GB
GGUF: No

Sparse activation for efficiency; Strong on Chinese tasks

ERNIE 4.0-8B

Baidu · China

Restricted

Params: 8B
Context: 8,192
Min VRAM: -1 GB
GGUF: No

Integrated with Baidu ecosystem; Good Chinese NLP

ERNIE Speed-7B

Baidu · China

Restricted

Params: 7B
Context: 8,192
Min VRAM: -1 GB
GGUF: No

Fast inference for Baidu services; Optimized for speed

ERNIE Lite-3B

Baidu · China

Open

Params: 3B
Context: 4,096
Min VRAM: 3 GB
GGUF: Yes

Lightweight, partially open; Good for basic tasks

Kimi-K1.5-32B

Moonshot AI · China

Restricted

Params: 32B
Context: 131,072
Min VRAM: 24 GB
GGUF: No

Very long context; Strong reasoning, similar to GPT-4

Kimi-VL-7B

Moonshot AI · China

Restricted

Params: 7B
Context: 65,536
Min VRAM: 8 GB
GGUF: No

Multimodal with long context; Good image reasoning

Step-1-100B

StepFun · China

Restricted

Params: 100B
Context: 32,768
Min VRAM: -1 GB
GGUF: No

Large scale, strong general performance; Good for complex tasks

Step-2-7B

StepFun · China

Open

Params: 7B
Context: 8,192
Min VRAM: 6 GB
GGUF: Yes

Lightweight, open weights; Good for practical use

SenseNova-7B

SenseTime · China

Restricted

Params: 7B
Context: 8,192
Min VRAM: -1 GB
GGUF: No

Integrated with SenseTime vision models; Good in Chinese contexts

CodeGeeX4-9B

Zhipu AI · China

Open

Params: 9B
Context: 8,192
Min VRAM: 8 GB
GGUF: Yes

Multilingual code generation; Good completion capabilities

CPM-Bee-10B

OpenBMB (Tsinghua) · China

Open

Params: 10B
Context: 4,096
Min VRAM: 8 GB
GGUF: Yes

Bilingual optimization; Good for Chinese tasks

LongCat-2.0

Meituan · China

Open

Params: 1600B
Context: 1,000,000
Min VRAM: ,
GGUF: No

Near-frontier agentic coding; 1M-token context; official FP8 and INT8 quant repos on Hugging Face (created 2026-07-03 and 2026-07-05); topped OpenRouter as the anonymous 'Owl Alpha' before Meituan open-sourced it on 2026-06-30

Kimi K3

Moonshot AI · China

Open

Params: 2800B
Context: 1,048,576
Min VRAM: ,
GGUF: No

1M-token context; native vision; always-on reasoning; quantization-aware training with MXFP4 weights and MXFP8 activations; #1 on LMArena Frontend Code (1679 Elo) at the 2026-07-16 launch; API $3.00/M input (cache miss), $0.30/M cache hit, $15.00/M output

Notable vendors

AlibabaDeepSeekZhipuBaiduTencentMoonshot

Run benchmarks

llama.cpp, MLPerf, and inference leaderboards.

VRAM calculator

Will this model fit on your GPU?

Hardware guides

Local LLM setup, homelab, mini-PC builds.

Methodology

How we benchmark, verify ASINs, and source data.

Beginner's guide

New to local AI? Start here.

Sourcing & commerce disclosure

Model metadata on this page is compiled from each vendor's public HuggingFace repo, model card, or research paper at time of indexing. Performance figures (KMMLU, MMLU, HAERAE, AraBench, IndicGEM, etc.) link directly to source benchmarks where available; missing or partial values are shown as ", ".

Outbound hardware links (GPUs, mini-PCs, accelerators) are Amazon Associate links using store ID fredoline-20. As an Amazon Associate we earn from qualifying purchases. Prices and availability shown elsewhere on the site are snapshots, not live quotes — verify on Amazon before purchasing. See About / disclosures for the full policy.