amd/Kimi-K2.6-MXFP4

Model Overview

• Model Architecture: Kimi-K2.6
  • Input: Text, Image
  • Output: Text
• Supported Hardware Microarchitecture: AMD MI350/MI355
• ROCm: 7.0
• Operating System(s): Linux
• Inference Engine: vLLM
• Model Optimizer: AMD-Quark
  • Quantized layers: experts, shared_experts
  • Weight quantization: OCP MXFP4, Static
  • Activation quantization: OCP MXFP4, Dynamic
• Calibration Dataset: Pile

This model was built with Kimi-K2.6 model by applying AMD-Quark for MXFP4 quantization.

Model Quantization

The model was quantized from a BF16-decompressed version of moonshotai/Kimi-K2.6 using AMD-Quark. The original checkpoint uses native INT4 (compressed-tensors) quantization; it was first decompressed to BF16 before applying MXFP4 quantization. The weights and activations are quantized to MXFP4.

Quantization scripts:

cd Quark/examples/torch/language_modeling/llm_ptq/
exclude_layers="*self_attn* *mlp.gate *mlp.gate.linear *lm_head *mlp.gate_proj *mlp.up_proj *mlp.down_proj *mm_projector* *vision_tower*"

python quantize_quark.py \
    --model_dir /path/to/Kimi-K2.6-bf16 \
    --quant_scheme mxfp4 \
    --exclude_layers $exclude_layers \
    --output_dir amd/Kimi-K2.6-MXFP4 \
    --model_export hf_format \
    --file2file_quantization

Deployment

Use with vLLM

This model can be deployed efficiently using the vLLM backend.

Evaluation

The model was evaluated on GSM8K benchmarks.

Accuracy

Benchmark	Kimi-K2.6	Kimi-K2.6-MXFP4 (this model)	Recovery
GSM8K (flexible-extract)	0.9393	0.9318	99.2%

Reproduction

The GSM8K results were obtained using the lm-evaluation-harness framework, with vLLM, lm-eval and amd-quark compiled and installed from source.

lm_eval \
  --model vllm \
  --model_args pretrained=amd/Kimi-K2.6-MXFP4,trust_remote_code=True,tensor_parallel_size=4 \
  --tasks gsm8k \
  --batch_size auto

License

Modifications Copyright(c) 2026 Advanced Micro Devices, Inc. All rights reserved.