General and specialized, open-source and closed-source, text and multimodal — a one-stop catalog of LLMs from Chinese vendors, with benchmark rankings, API pricing, and deployment hardware references.
Organized by model capability — click to enter the corresponding group in the model library
Flagship closed-source and open-source MoE models with deep thinking, tool calling, and long context
28 models → 02Image / video understanding, visual reasoning, and screen-operating agents
8 models → 03Text-to-image and image editing, with 2K–3K high resolution and Chinese text rendering
10 models → 04Text/image-to-video, native audio, 2K resolution, and long-form storytelling
13 models → 05Speech recognition, speech synthesis, and real-time voice conversation LLMs
6 models → 06Open-source code models and coding assistants at the frontier of agentic programming
5 models → 07Vertical scenarios including finance, healthcare, education, government, and autonomous driving
7 models → 08Supporting capabilities such as vector representations, reranking, and web search
4 models →New models and major release news (October 2026)
Switchable thinking / non-thinking modes, 1M context, unchanged APIs and pricing, with weights also released; the same-generation V4-Pro-0813 is a 1.6T total-parameter MoE flagship (MIT open source).
The K3 series features 2.8T total parameters, Kimi Delta Attention architecture, and native multimodality; K3.1 Pro strengthens coding and long-horizon agents, with a HighSpeed variant offering roughly 3× throughput.
Z.ai's GLM-5 series has iterated to 5.3: the flagship targets complex software engineering and long-horizon agents; Flash / FlashX offer native multimodality in cost-effective and high-speed tiers, all with 1M context.
Built for long-horizon agent work, coding, and complex reasoning; the open-source Qwen3.5 / 3.6 / 3.8 lines continue to iterate, with the Qwen open-source ecosystem leading globally in downloads.
The new Hunyuan preview is available in open weights, targeting coding, agent tool calling, and long contexts; the current Hy3 offers a cost-effective API tier at $0.10/$0.41.
Closed and open models compete head-to-head: DeepSeek and Kimi stand in the first tier with open weights, while GLM, Qwen, andDoubao lead in coding and agent scenarios via API flagships.
A strong lineup with distinct strengths: MiniMax for long context and video, Tencent Hunyuan for the WeChat ecosystem, iFLYTEK trained on fully domestic compute, and StepFun for multimodal MoE.
Differentiated routes including on-device, academic, and search-augmented models, serving niche scenario needs.
Two things to check when choosing: how much it costs and what it runs on
DeepSeek's official peak/off-peak pricing plus list-price snapshots from all vendors (2026-10) — per-million-token costs at a glance.
View prices → ⚙Domestic chip roadmaps from Ascend / Cambricon / Hygon / Moore Threads, plus all-in-one machines and supernode solutions.
View hardware → ⚖Benchmarks and prices from the same sources for domestic flagships vs. GPT-6 / Claude / Gemini — strengths and gaps at a glance.
See comparison →