Side-by-side comparison of Gemma 4 26B A4B Instruct (Google DeepMind · United States) and Qwen3-8B (Alibaba Cloud (Qwen) · China) for self-hosted deployment of the open-weight model. Gemma 4 26B A4B Instruct is rated conditional; Qwen3-8B is conditional. They part ways on training data: Gemma 4 26B A4B Instruct is "Domain-level summary", Qwen3-8B is "Token count only".
| Field | ||
|---|---|---|
| Summary | ||
| Verdict | Conditional Based on published licence terms, Gemma 4 26B A4B ships under pure Apache 2.0 with no prohibited-use carve-outs — a departure from prior Gemma generations. The sparse-MoE architecture (25.2B total / 3.8B active) puts it in an ambiguous zone for EU AI Act GPAI systemic-risk classification, and US origin plus image-input support add transparency obligations that deployers should document. | Conditional Based on published licence terms, Qwen3-8B is released under standard Apache 2.0 with no field-of-use carve-outs, making self-hosted commercial deployment viable. Training-data disclosure is limited to a token count and Chinese origin creates EU AI Act Art. 53 transparency and data-transfer risks that deployers should document. |
| Last reviewed | 2026-04-17 | 2026-04-17 |
| Open-weight | ||
| Licence | Apache 2.0 | Apache 2.0 |
| Commercial use | Unrestricted | Unrestricted |
| Training data | Domain-level summary | Token count only |
| Origin | United States | China (Hangzhou) |
| Performance & pricing? | ||
| Quality index | 27/100 | 11/100 |
| Speed | — | 86 tok/s |
| Blended price | — | $0.31/M |
| Context window | — | — |
| Evidence | ||
| Sources | ||
No overlapping sources between the two entries.