What this is
The 2026 Chinese open-source LLM ecosystem is led by ~10 labs, all shipping flagship models at sub-quarter cadence. This page is a 2026-05-15 leaderboard reading on each lab's flagship model, scale, licensing posture, and standout strength.
Top Chinese Open-Source LLM Companies (2026)
| Lab | Flagship 2026 | Scale | License posture | Standout strength |
|---|---|---|---|---|
| DeepSeek | DeepSeek V4-Pro | 1.6T / 49B active | MIT | Coding (83.7% SWE-Bench), low cost |
| Alibaba (Qwen) | Qwen 3.6-Max Preview | Multi-size (0.6B-235B MoE) | Apache 2.0 open tier; closed Plus tier | Broadest size + modality coverage |
| Moonshot AI (Kimi) | Kimi K2.6 | 1T / 32B active | Open weights | Best open coder (SWE-Bench Pro 58.6) |
| Tencent (Hunyuan) | Hunyuan 3.0 / Hy3preview | 295-406B MoE | Open-source (most variants) | Modality breadth: text, video, 3D worlds |
| Z.ai / Zhipu (GLM) | GLM-5 / GLM-5.1 | 744B / 40B active | Open-source (5.1) | Public-company; quarterly cadence |
| Baidu (ERNIE) | ERNIE 5.0 | Proprietary at flagship; 4.5 open | Mixed | LMArena top-10; 1.5B daily API calls |
| ByteDance (Doubao / Seed) | Seed 2.0 | Multi-tier proprietary | Proprietary; lowest pricing | $0.022/M output floor |
| 01.AI (Yi) | Yi-Lightning | Mixed (open base, closed flagship) | Mixed | Pioneer; Yi-Coder still strong |
| Huawei (Pangu) | Pangu Ultra MoE | 718B | openPangu 7B + 72B open | Only frontier model trained on Chinese stack |
| MiniMax | abab series | Multi-tier proprietary | Proprietary | Long-context multimodal; HKEX listed 2026 |
Scale Signals (Q1 2026)
| Lab | Signal |
|---|---|
| DeepSeek | ~50M+ downloads; setting cost floor for the category |
| Alibaba Qwen | 11+ flagship releases in 18 months; broadest size range |
| Moonshot Kimi | Open-weight SWE-Bench Pro leader; first to beat GPT-5.4 |
| Tencent Hunyuan | Open-source video + world models; only lab with that breadth |
| Z.ai GLM | HKEX IPO Jan 2026; first publicly listed Chinese AI lab |
| Baidu ERNIE | 1.5B daily API calls; global LMArena top 10 |
| ByteDance Doubao | Set the price floor at $0.022/M output |
| Huawei Pangu | Frontier model on Chinese hardware stack only |
| MiniMax | IPO followed Z.ai on HKEX |
Six Things the Leaderboard Tells You
- The Chinese open-source landscape is denser than the Western one. Ten labs at frontier or near-frontier capability vs three or four in the West (Meta, Mistral, AI2, smaller).
- Open vs proprietary splits are converging on a pattern. Most Chinese labs open the base / mid tier and close the flagship + plus tier. DeepSeek (fully MIT) is the outlier.
- 2026 is the IPO year for Chinese AI labs. Z.ai (Jan 8) and MiniMax both went public on HKEX, changing capital structure and disclosure cadence.
- Specialisation by use case is sharpening. DeepSeek = coding + low cost; Kimi = agentic coding; Tencent = modality breadth; Z.ai = enterprise + Cambricon-compatible; Doubao = lowest cost.
- Huawei is unique on the hardware-independence axis. The only lab with frontier model + frontier training chip co-owned.
- Release cadence is now quarterly or faster. Most labs shipped 4-7 flagship-tier releases in the past 12 months.
What This Means for AI Visibility
Brand-visibility monitoring for the Chinese market in 2026 must cover at least DeepSeek, Qwen, Kimi, GLM, and ERNIE; the second tier (Hunyuan, Doubao, Pangu) covers specialised deployments. Western-only AI visibility monitoring misses a growing share of agentic surfaces, particularly inside agentic coding tools, multimodal applications, and Chinese consumer apps.
Methodology
Data sourced from each lab's release announcements and aggregated by the page-specific timeline pages on this site. See Qwen, Hunyuan, Kimi, ERNIE, Doubao, GLM, Yi, and Pangu.
How Presenc AI Helps
Presenc AI runs brand-visibility prompt suites against every major Chinese LLM (DeepSeek, Qwen, Kimi, GLM, ERNIE, Hunyuan, Doubao, Pangu) so brand teams selling into the Chinese market see how they appear across the actual deployment mix, not just the Western default surfaces.