new model · 2026.09.12
Running Ling-3.0-flash-VL locally: what the 70.87 GiB INT4 build means
Ling-3.0-flash-VL is an image/video-input MoE model with about 124.85B parameters. The SGLang guide lists roughly 5.1B active parameters, which does not describe the stored weight size. The official INT4 file is about 70.87 GiB. Published single-DGX-Spark measurements now let you compare output cadence across quantization formats.
Read 124.85B total and 5.1B active together
Ling-3.0-flash-VL is an MoE model that selects a subset of experts for computation. Its 124.85B figure describes the whole model, while 5.1B is the portion active for a token. The latter helps explain compute, but unselected expert weights still need to be stored and accessible during execution.
Do not size an ordinary laptop as though 5.1B were a small dense model. Beyond the files, the runtime needs workspaces and KV cache. Opening the full 256K context requires much more cache headroom than a short chat, so separate the advertised context ceiling from your normal input length.
BF16, INT4 and FP4 start with different file sizes
The files in the official repositories total about 232.55 GiB for BF16, 70.87 GiB for INT4 and 64.81 GiB for FP4. These are starting points for download and storage planning. Execution needs additional headroom, and conversion may require keeping source and output files at the same time.
The official FP4 label does not mean NVIDIA NVFP4 or GGUF, and INT4 is not automatically compatible with every engine's four-bit loader. Equal bit counts can still use different scaling layouts and kernels. Before downloading, identify the exact server image and recipe that supports that release.
Video input is different from video generation
This is a vision-language model that accepts text with images or video frames, interprets them and responds in text. It is not a model that creates new video. It may suit product-video summaries or questions about scenes, but video-generation comparisons belong to a different model family.
The SGLang configuration uses a native 128K context; 256K requires YaRN extension. Video frames and long documents share the input budget and memory, so start with one real task rather than maximizing frames, resolution and text length together.
How fast is it on one DGX Spark?
The official SGLang cookbook includes INT4 and MXFP4 measurements on one DGX Spark. With 8,192 input tokens, 1,024 output tokens and one concurrent request, median time per output token was 44.18 ms for INT4 and 30.18 ms for MXFP4, equivalent to roughly 22.6 and 33.1 decoded tok/s. MXFP4 and NVFP4 are different formats despite their similar names.
Under the same conditions, median time to first token was about 2.89 seconds for INT4 and 2.54 seconds for MXFP4. Our site separates prefill and decode using this 8K-input reference. Other input lengths are estimates, and these text measurements exclude image preprocessing. We do not transfer these rates to other GPUs or Macs.
Change log
These entries record changes to the site's guidance. They do not automatically check your installed engine or model version.
Single-Spark execution paths for new models
The speed experience now uses published single-Spark measurements for DeepSeek and Ling. The same speeds and commands are not applied to unverified hardware.