Ling-3.0
PulseAugur coverage of Ling-3.0 — every cluster mentioning Ling-3.0 across labs, papers, and developer communities, ranked by signal.
1 day(s) with sentiment data
-
Ant Group launches SingProbe, an in-process AI safety guardrail
Ant Group has introduced SingProbe, an internally developed safety system designed to integrate risk detection directly into the AI model's generation process. This approach aims to identify unsafe content and factual i…
-
ik_llama.cpp adds DFlash 2 speculative decoding and new model support
A pull request has been submitted to the ik_llama.cpp project, introducing DFlash 2 speculative decoding. This update also includes support for IQ4_KS and IQ4_KT quantization formats on RDNA3 GPUs via HIP, Vulkan, and i…
-
Open-weight AI transparency debated beyond downloadable endpoints
The discussion around open-weight AI models often centers on whether final weights are downloadable. However, a deeper question of transparency involves the ability for outsiders to inspect multiple endpoints or checkpo…
-
AntLing releases six Ling-3.0 base model checkpoints for research
AntLing has released the full set of six base checkpoints for its Ling-3.0 model, available in both tiny and flash sizes across three training stages: pretrained, mid-trained, and WSM-merged. These checkpoints are inten…
-
Ling-3.0 (BailingMoE3) model integrated into llama.cpp for local use
The Ling-3.0 (BailingMoE3) model has been officially integrated into the llama.cpp mainline, enabling its use in local setups. Quantized GGUF versions of the Ling-3.0 tiny (8B) and flash (127B) models are now available.…
-
Ling-3.0 model support added to llama.cpp via PR #26608
A pull request (PR #26608) has been submitted to the llama.cpp project to add support for the Ling-3.0 model architecture. This architecture is noted to be similar to DeepSeekV2. A small addition to support the Tiny mod…
-
inclusionAI releases lightweight Ling-3.0-tiny MoE model for local deployment
inclusionAI has released Ling-3.0-tiny, a new hybrid reasoning Mixture-of-Experts (MoE) model with 7.9 billion total parameters and 1.3 billion activated parameters per token. This model is designed for efficient local …