Qwen3.8-Flash-Next is a multimodal mixture-of-experts (MoE) model with a 125B-parameter main model supplemented by an additional 51B N-gram embeddings, with 6B parameters activated per token. It has a native 262,144-token context window, extensible to 1M tokens with YaRN
Performance on GB300 NVL72
Running the FP8 variant with TensorRT-LLM on GB300 NVL72 (72 Blackwell Ultra GPUs, 130 TB/s NVLink):
-
>16,000 tokens/sec per GPU
-
>200 tokens/sec per user
The model also runs on local NVIDIA hardware — DGX Spark clusters, DGX Station, and workstations with 4× RTX PRO 6000 Blackwell GPUs — and scales up to GB300 NVL72 for production.
Supported inference and training stacks
-
SGLang Cookbook
-
VLLM Recipe
-
TensorRT-LLM Day-0 Support
-
NeMo AutoModel SFT + LoRA fine-tuning
-
NeMo RL Reinforcement learning recipes
-
TokenSpeed Recipe
Get started
-
Weights: Hugging Face | ModelScope
-
Managed API (production version): QwenCloud
-
Full writeup: NVIDIA Technical Blog