Lookoff/gemma-4-26b-a4b-it-uncensored-4bit

Gemma 4 26B Uncensored — TurboFieldfare-Compatible 4-bit Checkpoint

Run an uncensored Gemma 4 26B-A4B model locally on Apple Silicon with approximately 2 GB of active RAM using TurboFieldfare Uncensored.

This is a purpose-built and validated 4-bit checkpoint for TurboFieldfare's Swift/Metal expert-streaming runtime. It uses the quantization recipe and tensor layout expected by TurboFieldfare: affine 4-bit weights with group size 64 and an 8-bit router projection.

The complete checkpoint occupies approximately 14.3 GB on disk. The approximately 2 GB RAM figure refers to the resident weights and KV cache used by TurboFieldfare while routed experts are streamed from SSD. Loading the complete checkpoint normally with MLX requires substantially more memory.

Why this checkpoint exists

TurboFieldfare does not simply load an arbitrary 4-bit model. Its hand-written Metal kernels and streaming installer expect a specific weight representation.

This checkpoint was produced and validated for:

• affine 4-bit quantization with group size 64;
• 4-bit embeddings, attention, and shared and routed expert weights;
• an 8-bit router.proj;
• the tensor structure consumed by TurboFieldfare's installer and Metal kernels;
• SSD-backed routed-expert streaming on Apple Silicon.

Generic MLX or GGUF quantizations should not be assumed to be drop-in TurboFieldfare inputs, even when they use the same base model and nominal bit width.

Quick start with TurboFieldfare

Requirements

• Apple Silicon Mac
• macOS 26 with Metal 4
• Xcode 26 and Swift 6.2 or newer
• approximately 15 GB available for download and 14.3 GB for the installed checkpoint

Clone and build the uncensored fork:

git clone https://github.com/Lookoff-AIMLAPI/turbo-fieldfare-uncensored.git
cd turbo-fieldfare-uncensored
swift build -c release

Install this checkpoint in TurboFieldfare's streaming format:

.build/release/TurboFieldfareRepack \
  --variant uncensored \
  --output scratch/gemma4-uncensored.gturbo

Run it with the CLI:

.build/release/TurboFieldfareCLI \
  --model scratch/gemma4-uncensored.gturbo \
  --messages-file msgs.json

Or launch the native Mac app with the uncensored variant selected:

TURBOFIELDFARE_VARIANT=uncensored .build/release/TurboFieldfareMac

See the TurboFieldfare Uncensored repository for the runtime, local OpenAI-compatible server, benchmarks, and complete usage documentation.

Use with plain MLX

The checkpoint can also be loaded directly with mlx-lm:

pip install -U mlx-lm
mlx_lm.chat --model Lookoff/gemma-4-26b-a4b-it-uncensored-4bit

Plain MLX loads the checkpoint normally and does not provide TurboFieldfare's approximately 2 GB active-RAM behavior.

Checkpoint details

Property	Value
Architecture	Gemma 4 26B-A4B Mixture of Experts
Behavior	Instruction-tuned, abliterated / reduced refusals
Format	MLX Safetensors
Weight quantization	Affine 4-bit
Group size	64
Router projection	8-bit
Installed size	Approximately 14.3 GB
TurboFieldfare active RAM	Approximately 2 GB of resident weights and 4K KV cache
Target runtime	TurboFieldfare on Apple Silicon
Secondary runtime	mlx-lm with normal full-checkpoint memory behavior

The RAM figure is runtime-specific. Prompt length, selected context size, KV-cache configuration, expert-cache settings, page-cache state, and macOS memory pressure affect observed memory use and performance.

Provenance

1. Architecture and original weights: google/gemma-4-26b-a4b, distributed under the Gemma Terms of Use.
2. Instruction tuning and abliteration source: TrevorJS/gemma-4-26B-A4B-it-uncensored. Its model card reports biprojection plus Expert-Granular Abliteration, a 0.7% refusal rate, and KL divergence of 0.09 relative to the base checkpoint.
3. This repository: custom affine quantization and packaging matching the 4-bit/group-64 layout required by TurboFieldfare, with router.proj retained at 8-bit.

For structural compatibility, the quantization layout follows the official mlx-community/gemma-4-26b-a4b-it-4bit release while applying it to the uncensored source checkpoint.

Scope and limitations

• This repository contains model weights, not the TurboFieldfare runtime itself.
• TurboFieldfare's current path is text-only; images, audio, and video are not supported.
• Reduced refusal behavior does not guarantee correctness, neutrality, or suitability for a particular use.
• Generated output may be inaccurate, harmful, or otherwise inappropriate. Evaluate important outputs independently.
• Performance varies by Mac, SSD, prompt, cache state, and runtime settings.

License and responsible use

The checkpoint inherits the Gemma license and remains governed by the Gemma Terms of Use and Gemma Prohibited Use Policy.

The model has reduced refusal behavior compared with the aligned instruction checkpoint and may produce content that the original model would decline. Users are responsible for complying with applicable law, the model terms, and the prohibited-use policy.

TurboFieldfare is an independent project and is not affiliated with, sponsored by, or endorsed by Google.