Release of Ling-3.0 Model Series and Base Checkpoints
InclusionAI has released the Ling-3.0 model series, comprising two primary versions: Ling-3.0-tiny (8B) and Ling-3.0-flash (127B). Alongside these post-trained models, the lab has released several base checkpoints on HuggingFace, including pretrained, mid-trained, and merged checkpoints. These base models are intended to support continued pretraining, fine-tuning, and academic research. According to the developer, the checkpoints represent different stages of the training process, ranging from large-scale pretraining to stages involving WSM merging.
Technical Specifications and Performance Benchmarks
The Ling-3.0 models utilize a Mixture-of-Experts (MoE) architecture. The “tiny” variant features approximately 7.89 billion total parameters, with 1.3 billion active parameters. The “flash” variant is significantly larger, reported by users on Reddit to be a 124B-A5.1B model.
Early performance data shared by users on Reddit suggests high efficiency for the “tiny” model. In local testing on an Intel Arc B580 GPU using the Vulkan backend, the Ling-3.0-tiny Q8_0 quantization achieved token generation speeds of approximately 114 tokens per second (t/s) and prompt processing speeds of 120.76 t/s. When tested on an NVIDIA GeForce RTX 3090 using CUDA, the model reached generation speeds of 186.91 t/s.
Some users have noted that the model’s reasoning capabilities, according to unverified benchmarks on Artificial Analysis, appear to exceed those of the Qwen3.5 9B model. However, other users have reported that the model may struggle with instruction following in specific tasks, such as translation, compared to other models like Gemma 4 E4B.
Software Support and Local Deployment
The Ling-3.0 series has officially gained support in llama.cpp as of build b10472+. This integration allows for local deployment via various quantization methods. Users have already released GGUF imatrix quantizations for both the tiny and flash models, facilitating easier use on consumer-grade hardware.
While the models show high performance in local environments, some users have raised questions regarding the stability of the model’s “thinking” processes, noting instances of looping behavior. Additionally, there is ongoing discussion regarding the efficacy of the model’s agentic and coding skills, with some users suggesting that the active parameter count may impact its performance in complex agentic workflows.