--- license: apache-2.0 license_link: https://ai.google.dev/gemma/docs/gemma_4_license thumbnail: https://huggingface.co/AtomicChat/gemma-4-26B-A4B-it-MLX-4bit/resolve/main/hero.png base_model: - google/gemma-4-26B-A4B-it base_model_relation: quantized quantized_by: AtomicChat pipeline_tag: text-generation library_name: mlx tags: - atomic-chat - gemma - gemma4 - google - mlx - apple-silicon - quantized ---

Base model: google/gemma-4-26B-A4B-it

**Gemma 4 26B A4B**, self-quantized to MLX by [Atomic Chat](https://atomic.chat). Built straight from Google's original weights with a per-tensor importance matrix, so this is not a repack of somebody else's files. Runs fully offline. ## Highlights - **25.2B total / 3.8B active per token parameters**: the weights this repo quantizes. - **Context length**: 256K tokens, as published by Google. - **30 layers**: Mixture-of-Experts, hybrid sliding-window (1024) and global attention. - **Modalities**: Text, Image. - **Full imatrix ladder**: every quant is calibrated with an importance matrix. - **Reasoning**: All models in the family are designed as highly capable reasoners, with configurable thinking modes. - **Diverse & Efficient Architectures**: Offers Dense and Mixture-of-Experts (MoE) variants of different sizes for scalable deployment. > [!NOTE] > These MLXs are **self-quantized from the original weights**, not a repack. The importance matrix keeps low-bit quants closer to the full-precision model. ## Model Overview | Property | Value | |---|---| | Base model | `google/gemma-4-26B-A4B-it` | | Parameters | 25.2B total / 3.8B active per token | | Layers | 30 | | Experts | 128 routed (top-8) | | Sliding window | 1024 tokens | | Context length | 256K tokens | | Vocabulary | 262K | | Modalities | Text, Image | | Architecture | Mixture-of-Experts, 128 experts (top-8), hybrid sliding-window (1024) and global attention, 16 attention heads over 8 KV heads, `Gemma4ForConditionalGeneration` | | This repo | MLX weights | ## Benchmarks | Benchmark | Score | |---|---| | MMLU Pro | 82.6% | | AIME 2026 no tools | 88.3% | | LiveCodeBench v6 | 77.1% | | Codeforces ELO | 1718 | | GPQA Diamond | 82.3% | | Tau2 (average over 3) | 68.2% | | HLE no tools | 8.7% | | HLE with search | 17.2% | | BigBench Extra Hard | 64.8% | | MMMLU | 86.3% | | MMMU Pro | 73.8% | | OmniDocBench 1.5 (average edit distance, lower is better) | 0.149 | | MATH-Vision | 82.4% | | MedXPertQA MM | 58.1% | | MRCR v2 8 needle 128k (average) | 44.1% | Scores are Google's published results for the base `google/gemma-4-26B-A4B-it`, not our own measurements. Quantization preserves the large majority of this; `Q4_K_M` and up stay close to full precision. ## Get started - **[Atomic Chat](https://atomic.chat):** search `AtomicChat/gemma-4-26B-A4B-it-MLX-4bit` and hit **Use this model**. - **mlx-lm:** `mlx_lm.generate --model AtomicChat/gemma-4-26B-A4B-it-MLX-4bit --prompt "Hello" --max-tokens 512` - **Server:** `mlx_lm.server --model AtomicChat/gemma-4-26B-A4B-it-MLX-4bit --port 8080` ## Best practices | Parameter | Value | |---|---| | temperature | 1.0 | | top_p | 0.95 | | top_k | 64 | Google's recommended sampling configuration for `google/gemma-4-26B-A4B-it`. ## How these were made 1. Download `google/gemma-4-26B-A4B-it` (original weights). 2. Convert and quantize with `mlx_lm.convert` on our pipeline. ## License Original model by Google, released under the Apache 2.0 license. Full terms: [Apache 2.0](https://ai.google.dev/gemma/docs/gemma_4_license). Quantized by Atomic Chat.