Ling 3.1 Flash 1019

Ling-3.1-flash-1019 is an open-weight hybrid reasoning model developed by InclusionAI, representing an upgraded iteration of the Ling-3.1-flash series. Built upon a hybrid-linear Mixture-of-Experts (MoE) architecture, the model scales to 560 billion total parameters while activating approximately 25 billion parameters per token. It features a native context window of 1 million tokens to accommodate long-document processing and extended agentic workflows.

Architecture and Specifications

The underlying architecture of Ling-3.1-flash-1019 integrates 63 Kimi Delta Attention (KDA) linear-attention layers interleaved with 9 gated Multi-head Latent Attention (MLA) layers at a 7-to-1 ratio, alongside 4 dense layers. Its sparse routing network configures 512 routed experts alongside a single shared expert, activating 8 experts per incoming token. The model maintains a hidden size of 4,096 with 64 attention heads and 64 KV heads, and supports multi-token prediction (MTP) for speculative decoding.

Capabilities and Performance

With post-training optimizations focused on complex reasoning and task execution, Ling-3.1-flash-1019 demonstrates enhancements across software engineering, coding, and general-purpose agent benchmarks. It shows notable gains on specialized evaluation suites such as AA-Briefcase and AutomationBench-AA compared to its predecessor checkpoint. The model is made available under the permissive MIT license, enabling flexible enterprise integration and community-driven fine-tuning.