Ling-3.0-flash open weights released in BF16 and FP8
- Source
- x.com
- Date
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.
Context
Ant Ling, the Ling model team at Ant Group, released the open weights for Ling-3.0-flash on August 4, 2026, after introducing the model on July 23. According to its model card, Ling-3.0-flash is a model, one that routes each through a few of its many specialist sub-networks, with 124 billion total parameters and 5.1 billion active for each token, released under the MIT license. The post offers two official versions so deployers can pick by hardware and performance needs. The BF16 version stores each weight as a 16-bit number. The FP8 version stores the expert weights as 8-bit numbers, in blocks of 128 by 128 values that each share a scale, which halves the bits those weights take.
The cost of the smaller format, in Ant's own testing, is a small accuracy loss. The FP8 model card reports GPQA-diamond at 84.00 against 84.97 for BF16, IFBench at 73.40 against 74.49, SciCode at 40.37 against 41.24, and ArcPrize at 67.18 against 68.75. Those four benchmarks are the only comparison the card gives, so the trade-off on other tasks is unmeasured, and the card gives no memory figures. For FP8 serving with SGLang it recommends four 141GB-class GPUs or a four-GPU Blackwell node, and eight GPUs on H100 or H800 machines. Ant followed this release the next day with 4-bit builds aimed at a single desktop machine.
- open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
- 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.
Checking sign-in…
Loading comments…






