zeta-2.1-mlx-mixed-3-4bit

Zeta 2.1 MLX — Mixed 3/4-bit

Mixed 3-bit / 4-bit MLX quantization of zed-industries/zeta-2.1, optimized for local edit prediction on Apple Silicon.

• Format: MLX
• Quantization: mixed 3-bit / 4-bit
• Use case: code edit prediction / autocomplete
• Tested on: Apple M4 Pro, 24 GB unified memory
• Client: Zed
• Server: MLX-LM

This is an independently produced quantized conversion and is not an official Zed Industries release.

Usage

Install MLX-LM:

uv tool install mlx-lm

Run the model:

HF_HUB_DISABLE_TELEMETRY=1 \
mlx_lm.server \
  --model stratorys/zeta-2.1-mlx-mixed-3-4bit \
  --host 127.0.0.1 \
  --port 9090 \
  --prefill-step-size 1024

Zed

Add to settings.json:

{
  "show_edit_predictions": true,
  "edit_predictions": {
    "provider": "open_ai_compatible_api",
    "mode": "eager",
    "open_ai_compatible_api": {
      "api_url": "http://127.0.0.1:9090/v1/completions",
      "model": "stratorys/zeta-2.1-mlx-mixed-3-4bit",
      "prompt_format": "zeta2_1",
      "max_output_tokens": 128,
      "prediction_debounce": 200
    }
  }
}

Offline usage

Once the model is cached:

HF_HUB_OFFLINE=1 \
HF_HUB_DISABLE_TELEMETRY=1 \
mlx_lm.server \
  --model stratorys/zeta-2.1-mlx-mixed-3-4bit \
  --host 127.0.0.1 \
  --port 9090 \
  --prefill-step-size 1024

Quantization

This repository contains a mixed 3-bit / 4-bit MLX quantization of Zeta 2.1.

The goal is to reduce model size, memory bandwidth, and inference cost while preserving edit-prediction quality as much as possible.

No additional fine-tuning was performed.

Limitations

Quantization may slightly reduce prediction quality compared with the original model.

Generated edits may be incorrect or fail to compile and should be reviewed before use.

Zeta 2.1 is specialized for edit prediction and is not intended as a general-purpose chat or agent model.

Attribution

Original model:

zed-industries/zeta-2.1

Upstream base model:

ByteDance-Seed/Seed-Coder-8B-Base

License

The upstream Zeta 2.1 model is distributed under the Apache License 2.0.

This repository is a quantized derivative. Please refer to the upstream model repository for the authoritative license and attribution requirements.