qwen3.8-Flash-Next#

  • Context Length: 262144

  • Model Name: qwen3.8-Flash-Next

  • Languages: en, zh

  • Abilities: chat, vision, tools, reasoning, hybrid

  • Description: Qwen3.8-Flash-Next is a native vision-language Mixture-of-Experts model with 125B main-model parameters, 6B activated per token, and an additional 51B n-gram embedding table. Its Qwen4 experimental architecture combines Gated DeltaNet, Qwen Sparse Attention, gated residuals, and flexible thinking control.

Specifications#

Model Spec 1 (pytorch, 125 Billion)#

  • Model Format: pytorch

  • Model Size (in billions): 125

  • Quantizations: none

  • Engines: vLLM, SGLang

  • Model ID: Qwen/Qwen3.8-Flash-Next

  • Model Hubs: Hugging Face, ModelScope

Execute the following command to launch the model, remember to replace ${quantization} with your chosen quantization method from the options listed above:

xinference launch --model-engine ${engine} --model-name qwen3.8-Flash-Next --size-in-billions 125 --model-format pytorch --quantization ${quantization}

Model Spec 2 (fp8, 125 Billion)#

  • Model Format: fp8

  • Model Size (in billions): 125

  • Quantizations: FP8

  • Engines: vLLM, SGLang

  • Model ID: Qwen/Qwen3.8-Flash-Next-FP8

  • Model Hubs: Hugging Face, ModelScope

Execute the following command to launch the model, remember to replace ${quantization} with your chosen quantization method from the options listed above:

xinference launch --model-engine ${engine} --model-name qwen3.8-Flash-Next --size-in-billions 125 --model-format fp8 --quantization ${quantization}

Model Spec 3 (fp4, 125 Billion)#

  • Model Format: fp4

  • Model Size (in billions): 125

  • Quantizations: FP4

  • Engines: vLLM, SGLang

  • Model ID: nvidia/Qwen3.8-Flash-Next-NVFP4

  • Model Hubs: Hugging Face, ModelScope

Execute the following command to launch the model, remember to replace ${quantization} with your chosen quantization method from the options listed above:

xinference launch --model-engine ${engine} --model-name qwen3.8-Flash-Next --size-in-billions 125 --model-format fp4 --quantization ${quantization}