Qwen3.8 Flash Next NVFP4 QAD released
- Source
- huggingface.co
- Author
- NVIDIA model releases
- Date

Context
NVIDIA has published an experimental, more heavily compressed version of Qwen's Qwen3.8-Flash-Next, a model that reads text, images and video and that Qwen released in August as a preview of its Qwen4 design. Compression here means : storing the model's weights in fewer bits so it takes less memory and computes faster. NVFP4 is NVIDIA's 4-bit number format and MXFP8 is an 8-bit format, and on NVIDIA's Blackwell GPUs the model computes in these formats directly. According to the model card, NVIDIA's earlier NVFP4 release quantized the routed experts, the specialist sub-networks a mixture-of-experts model picks between for each token. This version also stores the always-used shared experts in NVFP4 and selected attention projections in MXFP8. Its weight files take about 129.7 GB.
The added training is QAD, quantization-aware . A frozen full-precision copy of the original model acts as teacher, while the compressed student runs with simulated low-precision arithmetic and learns to match the teacher's outputs, giving it a chance to adjust to the rounding errors compression introduces. NVIDIA's write-up for its Nemotron models explains the same method at more length. The training here was short: this checkpoint was exported after 100 optimizer updates over 51,200 sequences, from a schedule set for 500.
NVIDIA's comparison with the original 16-bit model shows scores holding close rather than improving. GPQA Diamond moved from 92.05 to 91.60, IFBench from 79.60 to 79.93, SciCode from 50.22 to 51.11 and MMMU-Pro from 77.75 to 77.34, and Terminal-Bench 2.1 was tied at 85.96. The card says accuracy gains from QAD are not statistically established, that accounting for retries and timeouts in the runs is incomplete, and that the measurements do not establish an overall improvement over the original.
The release is also narrow in where it is known to work. NVIDIA validated it only on GB300 hardware with vLLM 0.30.0, four-way tensor parallelism and expert parallelism, plus helper code that refuses unvalidated builds. Generic loading, the model's built-in multi-token prediction module and accuracy at its maximum were not validated. NVIDIA describes the checkpoint as intended for quantization research and deployment validation. It is evidence that a more aggressive compression recipe can stay near the original's scores in NVIDIA's own tests, not a drop-in replacement for the earlier release.
- 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.
- 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.
- distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
Checking sign-in…
Loading comments…





