Ling 2.6 1T
#72 en Modèles de langage open sourceAnt Group · v2.6 · 1t · 7× · vu le 30 juin 2026
Ling-3.0-flash is an Open-Weight language model developed by Ant Group's InclusionAI (Ling series) featuring a hybrid linear Mixture-of-Experts architecture: 124 billion total parameters with approximately 5.1 billion activated per Token (1/64 Sparse-MoE-Routing combined with Kimi Delta Attention and Gated MLA). It supports a native context window of 262,144 Token (expansion to 1M Token announced), is text-based (no multimodal input), and optimized for agentic applications such as coding, tool calls, and long-context tasks. The model was released around July 23, 2026, with weights available via Hugging Face among other sources; independent Benchmark evaluations (e.g., Artificial Analysis) became available shortly after release.
Fonctionnalités
| Key-Benchmark (%) | GPQA Diamond: 59,3%; Artificial Analysis Intelligence Index: 26 |
| Kontextfenster (Token) | 262.144 Token nativ, Erweiterung auf bis zu 1M Token vorgesehen |
| Lizenz | MIT-Lizenz (laut abgeleitetem Modell auf Hugging Face, das die Lizenz vom Basismodell übernimmt) |
| Multimodalität | Nur Text (Input und Output), keine Bild-/Audio-Unterstützung |
| Plattform | Verfügbar über Hugging Face (Gewichte), API bei InclusionAI/Ant Group, OpenRouter, ZenMux, Kilo Code, Puter |
| Preis pro 1M Token | 0,07 $ Input / 0,22 $ Output pro 1M Token (InclusionAI-API, Stand Artificial Analysis) |
| Release-Datum | 23. Juli 2026 (laut OpenRouter-Modell-ID-Erstellung); Artificial Analysis nennt 4. August 2026 |