mavis-ai/Gemma4-26B-QAT-Q4

This repository contains an MLX-ready QAT Q4 distribution of Google's Gemma 4 26B A4B instruction-tuned Mixture-of-Experts model, prepared for local inference on Apple Silicon Macs.

This is not a newly fine-tuned model. It is a quantized MLX redistribution converted from Google's official Quantization-Aware Training (QAT) unquantized checkpoint. No structural modifications or additional training have been applied to the model architecture.

Important Notice

This repository is hosted primarily as a dedicated engine source for the R.E.V.I.S. application ecosystem. You are free to download and use this model package for your own local MLX projects or workflows, subject to the Apache License 2.0 and Google's Gemma terms.

For the original model cards, architecture details, intended usage, limitations, and evaluation information, refer to the official upstream models:

• Base model: https://huggingface.co/google/gemma-4-26B-A4B-it
• Official QAT source checkpoint: https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-unquantized
• Official QAT Q4 GGUF reference: https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-gguf

Quantization

This package uses MLX affine quantization with a mixed layout chosen to follow Google's official QAT Q4 GGUF release as closely as practical in MLX:

• Source checkpoint: google/gemma-4-26B-A4B-it-qat-q4_0-unquantized
• Reference layout: google/gemma-4-26B-A4B-it-qat-q4_0-gguf
• Default language tensor quantization: MLX affine Q4
• Group size: 64
• Mode: affine
• Token embeddings: MLX affine Q6
• Shared MLP and expert tensors: MLX affine Q4
• Router projection and router scale tensors: BF16
• Vision and multimodal projection components: BF16
• Norms, biases, and small non-linear tensors: preserved in their source floating-point representation

In practical terms, this is not a plain "quantize every linear layer to Q4" package. The embedding table is intentionally kept at higher precision, and the MoE router is kept in BF16 rather than quantized down with the expert weights. This keeps the MLX package closer to the official Google QAT Q4 distribution while still reducing local storage and memory requirements.

Optimized for R.E.V.I.S. (Local Cognitive OS)

We host and test this model package to serve as a local reasoning and judgment engine for R.E.V.I.S.

R.E.V.I.S. is a 100% local Cognitive OS for Multi-Agentic AI. It transforms your Mac devices into a distributed Agentic Swarm via zero-config Wi-Fi clustering, allowing you to run heavy AI workloads like recursive web research, dynamic RAG generation, and multi-step logic without killing single-machine performance.

If you are interested in pushing the limits of local AI and open-weight models, check out our project.

• Official Website: https://mavis-ai.co.jp/revis/
• Watch the 13-min Raw Demo (Multi-node Dynamic RAG): https://x.gd/LxaBF
• Follow our updates on X: https://x.com/mavis_ai_jp

Usage

Install or update the MLX runtime you use for Gemma 4 / multimodal models:

pip install -U mlx mlx-lm mlx-vlm huggingface_hub hf_xet

Download the model:

hf download mavis-ai/Gemma4-26B-QAT-Q4 \
  --local-dir ~/Models/mlx/Gemma4-26B-QAT-Q4

Run a quick text generation test:

python -m mlx_vlm.generate \
  --model ~/Models/mlx/Gemma4-26B-QAT-Q4 \
  --max-tokens 256 \
  --temperature 0.0 \
  --prompt "日本語で短く自己紹介してください。"

Usage Notes

Gemma 4 26B A4B is a Mixture-of-Experts model. Although only a subset of parameters is active per token, the full model still needs to be stored and loaded. On Apple Silicon, practical speed and context length depend heavily on unified memory size, MLX / mlx-vlm versions, prompt length, KV cache usage, and generation settings.

For day-to-day local RAG, a smaller Gemma 4 E2B / E4B QAT Q4 model may be faster. This 26B MoE QAT Q4 build is best suited for heavier judgment tasks, long-document inspection, extraction-quality review, and second-opinion reasoning.

Files

Recommended repository files:

README.md
LICENSE
config.json
generation_config.json
model-00001-of-00003.safetensors
model-00002-of-00003.safetensors
model-00003-of-00003.safetensors
model.safetensors.index.json
processor_config.json
tokenizer.json
tokenizer_config.json
chat_template.jinja

License

This repository redistributes a quantized derivative of Google's Gemma 4 26B A4B instruction-tuned QAT release, which is distributed by Google under the Apache License 2.0.

This derivative is likewise distributed under the Apache License 2.0. A copy of the license is included in the LICENSE file in this repository, and can also be found at https://www.apache.org/licenses/LICENSE-2.0.

Modification Notice

Compared with the official Google QAT source checkpoint, this repository applies the following packaging modification:

The source checkpoint was converted to MLX format and quantized with a GGUF-like mixed Q4/Q6/BF16 policy for local MLX inference.

No fine-tuning, additional training, or architecture-level modification has been applied.