Thesis
Qwen3 (Apache 2.0) and DeepSeek-R1/V3 (MIT) are the two strongest open-source Chinese LLM routes in 2026 — and both are globally available without restrictions. Qwen3-235B-A22B (April 2026) is the most capable open-weight model in either Chinese or Western labs as of May 2026 on MMLU (88.7) and MATH (79.4). DeepSeek-V3 (MIT, December 2024) leads on code (HumanEval 84.2). The practical difference between the two license tracks is minimal: Apache 2.0 adds a patent grant clause that enterprise legal teams often prefer; MIT is simpler. Both allow commercial fine-tuning, hosting-as-a-service, and distribution of derivatives. The harder questions are infrastructure (what GPU can run the model) and tooling (does the fine-tuning framework support it) — not license.
Decision in 20 seconds
| Use case | Best open-source option | License |
|---|---|---|
| Commercial fine-tuning (general purpose) | Qwen3-7B or Qwen3-14B | Apache 2.0 |
| Fine-tuning for reasoning / STEM | DeepSeek-R1-Distill-Qwen-7B | MIT |
| On-prem inference, single A100 40GB | Qwen3-30B-A3B MoE (full precision) or Qwen3-32B (FP8) | Apache 2.0 |
| Local inference, M-series Mac | Qwen3-8B-Q4_K_M (Ollama) or DeepSeek-R1-7B (Ollama) | Apache 2.0 / MIT |
| Bilingual Chinese-English fine-tuning | GLM-4-9B or InternLM2-7B | Apache 2.0 |
Open source Chinese AI models — full list (2026)
| Model | Lab | Release date | License | Parameters | Download | Commercial use restrictions |
|---|---|---|---|---|---|---|
| Qwen3 (0.6B–235B) | Alibaba | April 2026 | Apache 2.0 | 0.6B / 1.7B / 4B / 8B / 14B / 30B-A3B / 32B / 235B-A22B | huggingface.co/Qwen | None (attribution only) |
| DeepSeek-V3 | DeepSeek | December 2024 | MIT | 671B total / 37B active (MoE) | huggingface.co/deepseek-ai | None (copyright notice only) |
| DeepSeek-R1 | DeepSeek | January 2025 | MIT | 671B total; distill variants: 1.5B–70B | huggingface.co/deepseek-ai | None (copyright notice only) |
| MiniMax-Text-01 | MiniMax | January 2025 | MIT | 456B total / 45.9B active (MoE) | huggingface.co/MiniMaxAI | None (copyright notice only) |
| GLM-4-9B | Zhipu AI (THUDM) | June 2024 | Apache 2.0 | 9B | huggingface.co/THUDM | None (attribution only) |
| InternLM2-7B / 20B | Shanghai AI Lab | January 2024 | Apache 2.0 | 7B / 20B | huggingface.co/internlm | None (attribution only) |
| Baichuan2-7B / 13B | Baichuan AI | September 2023 | Custom (Baichuan) | 7B / 13B | huggingface.co/baichuan-inc | Free commercial use below 100M MAU; custom license required above threshold |
| CogVideoX-2B / 5B | Zhipu AI (THUDM) | August 2024 | Apache 2.0 | 2B / 5B (video generation) | huggingface.co/THUDM | None (attribution only; video gen model) |
License comparison — what matters for your product
| License type | Models using it | Can fine-tune commercially? | Can host as a service (SaaS)? | Key requirement | Enterprise legal notes |
|---|---|---|---|---|---|
| Apache 2.0 | Qwen3, GLM-4-9B, InternLM2, CogVideoX | Yes, without restriction | Yes, without restriction | Preserve copyright notice + NOTICE file if present | Patent grant clause protects licensee from patent claims by licensor — preferred by many enterprise legal teams over MIT |
| MIT | DeepSeek-V3, DeepSeek-R1, MiniMax-Text-01 | Yes, without restriction | Yes, without restriction | Preserve copyright notice + MIT license text in distributions | Shorter and simpler than Apache 2.0; no explicit patent grant (slight theoretical risk); faster legal review cycle |
| Custom (Baichuan) | Baichuan2-7B/13B | Yes, below 100M MAU | Yes, below 100M MAU | Contact Baichuan AI for commercial license if MAU exceeds 100M | Practical blocker for large consumer apps; fine for enterprise internal tools or smaller SaaS products |
Why license routing matters for open-source Chinese models
The license landscape for Chinese open-source AI improved dramatically between 2023 and 2026. In 2023, most major Chinese models used custom licenses with vague commercial restrictions — LLaMA-style constraints that made enterprise deployment legally uncertain. The shift to Apache 2.0 (Qwen family from Qwen1.5 onwards) and MIT (DeepSeek from V2 onwards) resolved this entirely for the two strongest model families. The practical result: a team building a commercial product in 2026 has two fully permissive routes with state-of-the-art capability, neither requiring negotiation or legal review beyond standard OSS process. The main remaining exception is Baichuan2 — useful at smaller scales but the MAU cap makes it unsuitable for growth-stage consumer products.
Infrastructure is the more meaningful constraint now. Qwen3-235B-A22B at full precision requires 4×H100 for inference. The MoE architecture (only 22B parameters active per forward pass) makes it much more efficient than a dense 235B model, but it still requires significant hardware. For teams without that infrastructure: Qwen3-14B on a single A100 40GB delivers strong capability; Qwen3-30B-A3B MoE is the best quality-per-GPU-cost option at single-node scale.
FAQ
- What Chinese AI models are open source and commercially usable?
- Qwen3 (all sizes, Apache 2.0, Alibaba), DeepSeek-V3 and R1 (MIT), MiniMax-Text-01 (MIT), GLM-4-9B (Apache 2.0), InternLM2 (Apache 2.0). Baichuan2 is commercial below 100M MAU. ERNIE 4.0 and Kimi are proprietary API-only.
- Is Qwen3 Apache 2.0 licensed for commercial use?
- Yes — all Qwen3 sizes (0.6B through 235B-A22B MoE) are Apache 2.0 with no MAU cap, no revenue threshold, and no notification requirement. Attribution (preserving the license notice) is the only obligation. License text at github.com/QwenLM/Qwen3.
- What open source Chinese LLMs can I use for commercial fine-tuning?
- Qwen3-7B/14B (Apache 2.0) for general tasks; DeepSeek-R1-Distill-Qwen-7B (MIT) for reasoning-heavy tasks; InternLM2-7B (Apache 2.0) for Chinese scientific/technical text; GLM-4-9B (Apache 2.0) for bilingual fine-tuning. All support LLaMA-Factory, Axolotl, or Unsloth.
- How does DeepSeek-R1 MIT license compare to Qwen Apache 2.0?
- Both fully permit commercial use, fine-tuning, and hosting as a service. MIT is simpler (faster legal review). Apache 2.0 adds an explicit patent grant (protects licensee from patent claims). In practice: pick the model with better capability for your task — the license difference is rarely the deciding factor.
- Are there any restrictions on using Chinese open-source AI models globally?
- Qwen3 (Apache 2.0) and DeepSeek (MIT) impose no geographic restrictions. US export control law governs your downstream application but not the model weights themselves. Baichuan2 has an MAU cap in its custom license. Always verify the specific LICENSE file in each HuggingFace repo variant.
- Where can I download open source Chinese AI models?
- Qwen3: huggingface.co/Qwen. DeepSeek: huggingface.co/deepseek-ai. MiniMax-Text-01: huggingface.co/MiniMaxAI. GLM-4-9B: huggingface.co/THUDM. InternLM2: huggingface.co/internlm. China mirror: modelscope.cn. Quantized local inference: Ollama library (Qwen3 0.6B–14B, DeepSeek-R1 7B).
Companion pages in this cluster
| If your question is about… | Go to | What's there |
|---|---|---|
| Full benchmark and API comparison of China AI models | China AI Foundation Models | Qwen3, DeepSeek, Kimi, GLM, ERNIE — benchmarks, API access, use-case routing |
| Recent model releases and capability milestones | Model Release Tracker | Release timeline with benchmark data and verification sources |
| Weekly digest of what changed in China AI | China AI Updates | Weekly signal digest — model releases, funding, policy, curated for builders |
| Which China AI companies to watch (including open-source labs) | Best China AI Companies | 15-company shortlist: foundation model labs, workflow infra, physical AI |
| China AI startup funding and investment rounds | China AI Startup Funding Tracker | Moonshot $1B, MiniMax $600M, WeRide/Pony.ai IPOs — timeline with sources |
Quotable summary: In 2026, the license question for Chinese open-source AI is largely settled: Qwen3 (Apache 2.0) and DeepSeek (MIT) provide fully permissive commercial use for the two most capable open-weight model families in the world. The remaining decision is infrastructure — which GPU tier you have — and use-case fit: Qwen3 for multilingual and general fine-tuning, DeepSeek for code and reasoning. Everything else (Baichuan, GLM, InternLM) fills narrower niches with the same open-license principle.