Side-by-side comparison of Ling-2.6 1T (inclusionAI · China) and Qwen 3.5 (Alibaba · China) for self-hosted deployment of the open-weight model. Ling-2.6 1T is rated conditional; Qwen 3.5 is conditional. They part ways on licence: Ling-2.6 1T is "MIT", Qwen 3.5 is "Apache 2.0".
| Field | ||
|---|---|---|
| Summary | ||
| Verdict | Conditional Per the published model card, Ling-2.6 1T is an MIT-licensed 1-trillion-parameter MoE with a 262k-token context, hybrid MLA + Linear attention and multi-token-prediction support, targeted at production agentic workloads. Permissive weights enable EU self-hosting in principle, though the deployment footprint is non-trivial; vendor jurisdiction (Ant Group, China) and undisclosed training data remain the regulated-buyer blockers. | Conditional Hybrid Gated-DeltaNet + MoE flagship (397B total, 17B active) under Apache 2.0. Native vision, 201 languages, 262K context (1M with YaRN). Licence is clean; Chinese-origin alignment and supply-chain considerations persist. |
| Last reviewed | 2026-05-03 | 2026-04-15 |
| Open-weight | ||
| Licence | MIT | Apache 2.0 |
| Commercial use | Unrestricted | Yes |
| Training data | Undisclosed | Undisclosed |
| Origin | China | China |
| Performance & pricing? | ||
| Quality index | — | 40/100 |
| Speed | — | 53 tok/s |
| Blended price | — | $1.35/M |
| Context window | — | — |
| Evidence | ||
| Sources | ||
No overlapping sources between the two entries.