7.9B total / 1.3B active (MoE)
262,144 tokens with YaRN (131,072 native)
MIT
Compact native hybrid-reasoning model from Ant Group's inclusionAI, designed for local, edge, and lightweight agentic deployment with only 1.3B of its 7.9B parameters active per token. Ling-3.0-tiny uses a 3:1 hybrid-linear attention pattern comprising three Kimi Delta Attention layers followed by one Multi-Head Latent Attention layer in each four-layer block. Its sparse Mixture-of-Experts architecture contains 128 routed experts, activates eight routed experts plus one shared expert per token, and supports both fast non-thinking responses and multi-step reasoning through a per-request thinking control. The published deployment recipe extends its native 131,072-token context to 262,144 tokens with YaRN and includes structured tool calling, MTP/NEXTN speculative decoding, prefix caching, and Ling-specific SGLang and vLLM support. inclusionAI reports scores of 25 on Artificial Analysis Intelligence Index v4.1.1, 16 on the Artificial Analysis Agentic Index, 772 Elo on GDPval-AA v2, and 20.80 on tau3-Banking, with Artificial Analysis measuring output above 160 tokens per second. Local validation reports approximately 100-105 tokens per second on an NVIDIA DGX Spark and 86-90 tokens per second on an M4 Pro MacBook, using about 8.34 GiB of peak memory at an 8K context. The release provides official BF16, FP8, and INT4 checkpoints under the MIT license.