amd/Kimi-K2.6-NVFP4

Model Overview

• Model Architecture: Kimi-K2.6
  • Input: Text, Image, Video
  • Output: Text
• Supported Hardware Microarchitecture: AMD MI300/MI350/MI355 (emulation)
• ROCm: 7.2.2
• PyTorch: 2.10.0
• Transformers: 5.2.0
• Operating System(s): Linux
• Inference Engine: vLLM
• Model Optimizer: AMD-Quark (V0.12)
  • Quantized layers: experts and shared_experts
  • Weight quantization: NVFP4, Static
  • Activation quantization: NVFP4, Dynamic
• Calibration Dataset: Pile

This model was built with Kimi-K2.6 model by applying AMD-Quark for NVFP4 quantization.

Model Quantization

The model was quantized from moonshotai/Kimi-K2.6 using AMD-Quark. The weights and activations are quantized to NVFP4.

Quantization scripts:

cd Quark/examples/torch/language_modeling/llm_ptq/
export output_dir=amd/Kimi-K2.6-NVFP4
exclude_layers="*self_attn* *mlp.gate *mlp.gate.linear *lm_head *mlp.gate_proj *mlp.up_proj *mlp.down_proj *mm_projector* *vision_tower*"
python3 quantize_quark.py --model_dir $MODEL_DIR \
                          --quant_scheme nvfp4 \
                          --num_calib_data 128 \
                          --exclude_layers $exclude_layers \
                          --model_export hf_format \
                          --output_dir $output_dir \
                          --trust_remote_code \
                          --multi_gpu balanced 

Deployment

Use with vLLM

This model can be deployed efficiently using the vLLM backend.

Evaluation

The model was evaluated on GSM8K and MMLU_PRO benchmarks.

Accuracy

Benchmark	Kimi-K2.6	Kimi-K2.6-NVFP4(this model)	Recovery
GSM8K (flexible-extract)	93.93	93.48	99.52%
MMLU_PRO (exact-extract)	81.43	79.21	97.27%

Reproduction

The GSM8K and MMLU_PRO results were obtained using the lm-evaluation-harness framework, based on the Docker image rocm/vllm-dev:nightly_main_20260603.

Install the lm-eval (Version: 0.4.12) in container first.

pip install lm-eval
pip install lm-eval[api]

Launching server

export VLLM_ROCM_USE_AITER=1
vllm serve amd/Kimi-K2.6-NVFP4 -tp 8 \
  --mm-encoder-tp-mode data \
  --tool-call-parser kimi_k2 \
  --reasoning-parser kimi_k2 \
  --enforce-eager \
  --trust-remote-code

Evaluating model in a new terminal

lm_eval \
  --model local-completions \
  --model_args "model=amd/Kimi-K2.6-NVFP4,kv_cache_dtype=fp8,base_url=http://0.0.0.0:8000/v1/completions,tokenized_requests=False,tokenizer_backend=None,num_concurrent=32" \
  --tasks gsm8k \
  --num_fewshot 5 \
  --batch_size 1 

lm_eval \
  --model local-completions \
  --model_args "model=amd/Kimi-K2.6-NVFP4,kv_cache_dtype=fp8,base_url=http://0.0.0.0:8000/v1/completions,tokenized_requests=False,tokenizer_backend=None,num_concurrent=32,max_length=16384,timeout=14400" \
  --tasks mmlu_pro \
  --gen_kwargs "do_sample=True,temperature=1.0,top_p=0.95,max_tokens=4096,max_gen_toks=4096" \
  --batch_size auto \
  --limit 100

License

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