Kimi-K2.6-NVFP4

RedHatAI/Kimi-K2.6-NVFP4

Model Overview

• Model Architecture: moonshotai/Kimi-K2.6 (KimiK25ForConditionalGeneration)
• Input: Text, image, and video
• Output: Text
• Weight Quantization: NVFP4 (FP4 tensor-group quantization)
• Activation Quantization: NVFP4 (FP4 tensor-group quantization)
• Release Date: 2026-04-30
• Model Developers: RedHatAI

This model is a quantized variant of moonshotai/Kimi-K2.6, exported in compressed-tensors format for vLLM deployment and evaluated on instruction-following, reasoning, function-calling, and agentic coding workloads.

Model Optimizations

This checkpoint applies NVFP4 quantization to transformer linear layers with group-wise FP4 weights and activations, using FP8 scale tensors. The format is optimized for efficient low-precision serving while preserving strong benchmark quality on Kimi-K2.6 evaluations.

The model is exported in compressed-tensors format and is intended for OpenAI-compatible inference with vLLM.

Creation

This model was quantized with LLM Compressor and exported as compressed-tensors. The script below is a representative reference script aligned with recipe.yaml and the published quantization configuration.

<details> <summary><b>Reference quantization script (NVFP4)</b></summary>

from compressed_tensors.entrypoints.convert import CompressedTensorsDequantizer
from llmcompressor import model_free_ptq

MODEL_ID = "moonshotai/Kimi-K2.6"
SAVE_DIR = "Kimi-K2.6-NVFP4"

ignore = [
    "re:.*mlp.gate$",
    "re:.*lm_head",
    "re:.*self_attn.*",
    "re:.*kv_a_proj_with_mqa$",
    "re:.*q_a_proj$",
    "re:.*vision_tower.*",
    "re:.*embed_tokens$",
    "re:.*norm$",
    "re:.*mm_projector.*",
    "re:.*vision.*",
]

model_free_ptq(
    model_stub=MODEL_ID,
    save_directory=SAVE_DIR,
    scheme="NVFP4",
    ignore=ignore,
    converter=CompressedTensorsDequantizer(
        MODEL_ID,
        quant_config_key="text_config.quantization_config",
        ignore=ignore,
    ),
    max_workers=2,
    device="cuda:0",
)

</details>

Deployment

Use with vLLM

vllm serve RedHatAI/Kimi-K2.6-NVFP4 \
  --trust-remote-code \
  --mm-encoder-tp-mode data \
  --tool-call-parser kimi_k2 \
  --reasoning-parser kimi_k2 \
  --enable-auto-tool-choice

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="EMPTY")

resp = client.chat.completions.create(
    model="RedHatAI/Kimi-K2.6-NVFP4",
    messages=[{"role": "user", "content": "Explain how transformers use attention."}],
)

print(resp.choices[0].message.content)

Evaluation

We evaluated this model with lm-evaluation-harness, lighteval, BFCL v4, and SWE-Bench Lite served through a vLLM (0.22.1) OpenAI-compatible endpoint.

BFCL rows report category accuracy. SWE-Bench follows the official harness score style. For run transparency: 5 of 23 tasks were resolved, and 19 instances produced non-empty graded patches.

Recovery vs. base model (moonshotai/Kimi-K2.6)

Reproduction

Representative commands used to produce and aggregate these runs:

vLLM + lm-eval (example)

lm_eval --model local-chat-completions \
  --tasks gsm8k_platinum_cot_llama \
  --model_args "model=RedHatAI/Kimi-K2.6-NVFP4,max_length=40960,base_url=http://127.0.0.1:8000/v1/chat/completions,num_concurrent=32,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=3600" \
  --num_fewshot 0 \
  --apply_chat_template \
  --output_path results_gsm8k_platinum.json \
  --seed 1234 \
  --gen_kwargs "do_sample=True,temperature=1.0,top_p=0.95,top_k=20,min_p=0.0,max_gen_toks=64000,presence_penalty=1.5,repetition_penalty=1.0,seed=1234"

lighteval config used

model_parameters:
  provider: "hosted_vllm"
  model_name: "hosted_vllm/RedHatAI/Kimi-K2.6-NVFP4"
  base_url: "http://127.0.0.1:8000/v1"
  api_key: "EMPTY"
  timeout: 3600
  max_model_length: 40960
  concurrent_requests: 8
  generation_parameters:
    temperature: 1.0
    max_new_tokens: 65536
    top_p: 0.95
    seed: 1234
    top_k: 20
    presence_penalty: 1.5

lighteval endpoint litellm litellm_config.yaml \
  "aime25@1@8|0,math_500@1@3|0,gpqa:diamond@1@3|0" \
  --output-dir results_lighteval \
  --save-details

BFCL v4 and SWE-Bench Lite scripts

# BFCL categories: non_live, live, multi_turn, memory, web_search
./scripts/bfcl/run_bfcl_local.sh kimi_nvfp4 non_live
./scripts/bfcl/run_bfcl_local.sh kimi_nvfp4 live
./scripts/bfcl/run_bfcl_local.sh kimi_nvfp4 multi_turn
./scripts/bfcl/run_bfcl_local.sh kimi_nvfp4 memory
./scripts/bfcl/run_bfcl_local.sh kimi_nvfp4 web_search

# SWE-Bench Lite dev (full split)
SWEBENCH_SUBSET=lite SWEBENCH_SPLIT=dev SWEBENCH_SLICE= \
  ./scripts/swebench/run_swebench_lite_local.sh kimi_nvfp4

# Official SWE-bench resolved-rate evaluation
/home/shubhra/environments/mini-swe-agent/bin/python -m swebench.harness.run_evaluation \
  --dataset_name princeton-nlp/SWE-Bench_Lite \
  --split dev \
  --predictions_path /home/shubhra/kimik2.6_evals/results/swebench_resolved_eval/kimi_nvfp4_lite_dev_preds_merged.json \
  --max_workers 4 \
  --run_id kimi_nvfp4_lite_dev_20260701_resolved

Most lm-eval/lighteval tasks were run with 3 seeds and then averaged; AIME25 was run with 8 seeds. BFCL v4 and SWE-Bench Lite numbers come from the aggregated run artifacts listed below.

Every Eval Ever Artifacts

• every_eval_ever/aime25.json
• every_eval_ever/gpqa_diamond.json
• every_eval_ever/gsm8k_platinum_cot_llama.json
• every_eval_ever/ifeval.json
• every_eval_ever/math_500.json
• every_eval_ever/mmlu_pro_chat.json
• every_eval_ever/bfcl_v4.json
• every_eval_ever/swebench_lite_dev.json