Qwen3.5-9B-Base FP8
FP8_DYNAMIC conversion of Qwen/Qwen3.5-9B-Base at revision 68c46c4b3498877f3ef123c856ecfde50c39f404. This is the pretrained Base checkpoint, not the instruction-tuned model.
Linear weights use per-channel FP8 and activations use dynamic per-token FP8. Linear-attention projection layers are included. The vision tower, normalization layers, and language-model head retain their original precision. Tokenizer, processor, and context configuration come from the pinned source. No training or instruction tuning was performed.
See quantization-report.json for exact versions, exclusions, and a short-text loss comparison. That smoke check is not a comprehensive quality evaluation.