bouroo/zeta-2.1-mlx-mxfp4

A native MLX MXFP4 quantization of zed-industries/zeta-2.1, an 8B code edit-prediction / autocomplete model.

Quality note: Uniform MXFP4 quantization is provided as requested. In local mlx_lm smoke tests, the model loaded and ran but the canonical FIM prompt sometimes collapsed to repeated <|marker_1|> tokens. For reliable edit-prediction quality, use the MXFP8 variant or the original bf16 model. This artifact may still be useful for general code completion or experimentation.

Quantization details

Property	Value
Method	MLX MXFP4
Bits	4
Group size	32
Measured BPW	4.250
Size on disk	4.1 GB
Reference signal	bf16

MXFP4 is a block floating-point format supported by MLX on Apple Silicon. All eligible linear weights use MXFP4; the config records the quantization parameters.

About the base model

Zeta 2.1 is a code edit-prediction model (next-edit suggestion) finetuned from ByteDance-Seed/Seed-Coder-8B-Base. It uses an SPM-style, multi-marker FIM format rather than a chat template.

Prompt format

<[fim-suffix]>{code after cursor}
<[fim-prefix]><filename>{file_path}
{code before cursor}<|marker_1|><|marker_2|>
<[fim-middle]>

See examples/zeta_fim.py for a standard-library client targeting LM Studio.

Usage

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load("bouroo/zeta-2.1-mlx-mxfp4")
prompt = "<[fim-suffix]>\n    return total\n<[fim-prefix]><filename>app.py\ndef sum_list(items):\n<|marker_1|><|marker_2|>\n<[fim-middle]>"
out = generate(model, tokenizer, prompt=prompt, max_tokens=128,
               sampler=make_sampler(temp=0.2, top_p=0.95))

LM Studio

lms get https://huggingface.co/bouroo/zeta-2.1-mlx-mxfp4
lms load zeta-2.1-mlx-mxfp4

LM Studio's built-in autocomplete uses a simple FIM template and cannot format Zeta's multi-marker edit-prediction prompt. Use the client in examples/ or a compatible editor integration.

Verification

The model was converted with mlx_lm and confirmed loadable with local mlx_lm. See the quality note above for the observed uniform-MXFP4 FIM behavior.

---

Quantized and published under Apache-2.0, inherited from the base model.