Updates to Ant Ling’s Ling-3.1-flash Model
Ant Ling has announced a significant update to its Ling-3.1-flash model, introducing version 1019. According to an announcement on X (formerly Twitter), the update brings measurable improvements to the model’s performance across several key metrics.
The 1019 version has seen its ranking on the Artificial Analysis Intelligence Index rise to 44, up from a previous score of 41. Performance gains were also noted in agent evaluation benchmarks: the AutomationBench-AA score increased from 62% to 68%, while the AA-Briefcase Elo rating rose from approximately 1400 to 1524. While the update is now available, the developer has stated that API access and open weights are “coming soon,” though no specific release date has been provided.
Release of Ling-3.0-flash-Sante
In a separate announcement, InclusionAI has released a specialized version of the model titled Ling-3.0-flash-Sante. This model is built upon the Ling-3.0-flash architecture and features 124 billion total parameters, with 5.1 billion active parameters per token. It is designed with a 256K context window and is specifically optimized for medical reasoning, literature research, and tool-assisted workflows in the healthcare sector. Unlike the 3.1-flash update, the weights for the Sante model have already been made available on Hugging Face.
Local Performance Benchmarks on ARM Hardware
Independent testing shared on Reddit has provided unverified data regarding the ability to run large-scale models on consumer-grade ARM hardware. Using a setup consisting of two Lenovo IdeaCentre Mini units equipped with Snapdragon X processors (totaling approximately £600), a developer reported running the Qwen3.8 Flash Next model.
The testing utilized a custom, unreleased engine codenamed “ToRc,” which utilizes the CPU cores and the NPU for dense mathematical operations. According to the report, the setup achieved a read speed of 46.2 tokens per second (tok/s) and a generation speed of 20.3 tok/s when using the IQ4_NL quantization of the 95 GB model. The developer noted that because the model exceeds the 32 GB of RAM available in a single unit, the system must stream the “experts” from the NVMe storage.
While these results suggest that high-parameter models can achieve usable speeds on low-power hardware without a dedicated GPU, the findings remain anecdotal and have not been independently verified by third-party benchmarking firms.