gemma-4-26B-A4B-it-MNN

MNN HQQ 4-bit inference package converted from google/gemma-4-26B-A4B-it.

Original model

Gemma 4 26B A4B IT is Google DeepMind's instruction-tuned MoE vision-language variant. It accepts text and image input and produces text, with the upstream introduction emphasizing reasoning, coding and multilingual use. The original Gemma 4 family includes both dense and MoE architectures; this repository packages the 26B A4B checkpoint.

Summary adapted from the upstream model card. Original model capabilities and benchmark claims are separate from the MNN smoke checks below.

This checkpoint ties the token embedding to lm_head; llm_config.json records the shared weight offsets and tie_embeddings.quant_bit. The recorded --embed_bit 16 argument only controls a separately exported embedding and does not create BF16 token embedding weights here.

Download

python -m pip install -U huggingface_hub
python -c "from huggingface_hub import snapshot_download; snapshot_download('taobao-mnn/gemma-4-26B-A4B-it-MNN', local_dir='./gemma-4-26B-A4B-it-MNN')"
export MODEL_DIR="$(pwd)/gemma-4-26B-A4B-it-MNN"

MNN version and build

MNN: 7337404b2815b5fb9ddac13f4ac6a1fb37fcbbb4

This revision includes the MoE export fixes used for this release. Linux CPU build:

git clone https://github.com/alibaba/MNN.git
git -C MNN checkout 7337404b2815b5fb9ddac13f4ac6a1fb37fcbbb4
(cd MNN && ./schema/generate.sh)
cmake -S MNN -B MNN/build -DCMAKE_BUILD_TYPE=Release \
  -DMNN_BUILD_LLM=ON -DMNN_LOW_MEMORY=ON -DMNN_BUILD_CONVERTER=ON \
  -DMNN_SUPPORT_TRANSFORMER_FUSE=ON \
  -DMNN_BUILD_LLM_OMNI=ON -DMNN_BUILD_OPENCV=ON -DMNN_IMGCODECS=ON -DMNN_BUILD_AUDIO=ON
cmake --build MNN/build --parallel 4

Conversion

The quantization arguments match export_args.json. Download the original checkpoint and export it with the revision above:

python -m pip install -r MNN/transformers/llm/export/requirements.txt
python -c "from huggingface_hub import snapshot_download; snapshot_download('google/gemma-4-26B-A4B-it', local_dir='./original/gemma-4-26B-A4B-it', allow_patterns=['*.json', '*.safetensors', '*.model', '*.tiktoken', '*.jinja', '*.txt', '*.py'], ignore_patterns=['original/*', '*.gguf', '*.pt', '*.pth'])"

python MNN/transformers/llm/export/llmexport.py \
  --path ./original/gemma-4-26B-A4B-it --export mnn --hqq \
  --quant_bit 4 --quant_block 64 --scale_bit 16 --embed_bit 16 \
  --lm_quant_bit 4 --lm_quant_block 64 \
  --mnnconvert ./MNN/build/MNNConvert --dst_path ./reexported/gemma-4-26B-A4B-it-MNN

After re-export, reuse this package's verified tokenizer and template. Preserve the newly generated weight offsets and other model configuration:

python - <<'PY'
import json, os, shutil
from pathlib import Path
published = Path(os.environ["MODEL_DIR"])
converted = Path("./reexported/gemma-4-26B-A4B-it-MNN")
for filename in ("llm_config.json", "config.json"):
    path = converted / filename
    config = json.loads(path.read_text())
    reference = json.loads((published / filename).read_text())
    for key in ("jinja", "use_template"):
        config.pop(key, None)
        if key in reference:
            config[key] = reference[key]
    if filename == "config.json":
        config["tokenizer_file"] = "tokenizer.mtok"
    path.write_text(json.dumps(config, ensure_ascii=False, indent=2))
shutil.copyfile(published / "tokenizer.mtok", converted / "tokenizer.mtok")
PY

Run and test

printf 'What is 2 + 3? Reply with only the number.\n' > prompt.txt
./MNN/build/llm_demo "$MODEL_DIR/config.json" prompt.txt
./MNN/build/llm_bench -m "$MODEL_DIR/config.json"

The original migration passed 10 CPU High/Low smoke cases for this model. These are functional checks, not a broad quality or GPU benchmark.

License and attribution

Weights retain the upstream license and attribution. See the original model card, the license files in this repository and the MNN LLM documentation.