Today, we’re releasing the open weights for Ling-3.0-flash. 🎉
Official BF16 and FP8-quantized versions are now available, so you can choose the option that best fits your hardware, performance requirements, and deployment needs.
Get the Ling-3.0-flash weights here:
🤗 Hugging Face
Official
https://t.co/rCkfU23Li6
FP8
https://t.co/I6mZ1WhNQN
🤖 ModelScope
Official
https://t.co/hN7GJygYtG
FP8
https://t.co/h7e4OGeWo6
With the weights in your hands, you’re free to take Ling-3.0-flash further.
Run your own evaluations. Deploy it in your own environment. Adapt it to your use cases. Build it into your products, development toolchains, and agentic workflows.
Download it, make it yours, and show us what you build. 🚀
A huge thank-you to the @sgl_project and @vllm_project communities for their continued, high-quality support! Your contributions have helped make Ling-3.0-flash easier and more efficient to deploy across diverse hardware and production environments. 🚀
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
Why it matters
Self-hosting a 124B mixture-of-expertsA model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.Full definition → with 5.1B active parameters becomes possible: BF16 and FP8 weights on Hugging Face and ModelScope let you run private evaluations and deployments instead of depending on a provider endpoint.
