Xevn-4B-MLX-4bit

Transformers | MLX | GGUF | GitHub | Website

4-bit MLX weights of Xevn-4B for Apple silicon. Xevn is a 4B decision model built for Korean, trained for Korean understanding, knowledge, and judgment.

Xevn follows the Jev decision format. It reads a question and its options once and returns the answer together with a probability for every option. There is no text generation, so a request costs one forward pass.

Model overview

Type: Decision model with one forward pass per question and no text generation
Parameters: 4B
Architecture: Hybrid Gated DeltaNet and attention, 32 layers (24 Gated DeltaNet, 8 full attention)
Hidden size / vocabulary: 2,560 / 248,320
Question types: choice, noul, score
Languages: Korean (primary), English
Format: MLX
Quantization: 4-bit affine, group size 64
Size: 2.4 GB
Speed (p50): 78 ms

Quickstart

pip install -U mlx-lm huggingface_hub
hf download quevirus/Xevn-4B-MLX-4bit --local-dir Xevn-4B-MLX-4bit

import sys
sys.path.insert(0, "Xevn-4B-MLX-4bit")  # the repository ships the xevn_mlx.py helper
from xevn_mlx import XevnMLX

model = XevnMLX("Xevn-4B-MLX-4bit")

state = "고객: 어제 주문한 이어폰 값이 카드에서 두 번 결제됐어요. 오늘 안에 한 건 취소해 주세요."
# "Customer: the earphones I ordered yesterday were charged to my card twice. Please cancel one today."
questions = {
    "team": {"type": "choice", "instructions": "Which team should handle this ticket?",
             "criteria": {"billing": "Payments and refunds", "shipping": "Delivery", "account": "Login and account"}},
    "charged_twice": {"type": "noul", "instructions": "Was the customer charged more than once?"},
    "urgency": {"type": "score", "instructions": "How urgent is this ticket?",
                "criteria": ["Low", "Medium", "High"]},
}
result = {key: model.predict(state, q) for key, q in questions.items()}

Output (rounded):

{
  "team": {"answer": "billing", "probs": {"billing": 0.998, "shipping": 0.001, "account": 0.001}},
  "charged_twice": {"answer": true, "probs": {"yes": 0.993, "no": 0.007}},
  "urgency": {"answer": 1.77, "probs": {"0": 0.041, "1": 0.150, "2": 0.809}}
}

Values differ slightly between precisions.

Request format

A request is a state (text or any JSON value) and one or more named questions. Each question has a type, instructions, and criteria.

Type	Use	`criteria`	`answer` in the result
`choice`	Pick one of up to 26 options	`{"key": "description", ...}`, 2 to 26 options	The key of the most probable option
`noul`	Yes or no	Optional `{"true": "...", "false": "..."}`	`true` or `false`
`score`	Rate on a graded scale	`["lowest level", ..., "highest level"]`, 2 to 10 levels	The expected level (0 is the first level)

probs holds a probability for every option, yes/no, or level. For noul, a yes probability strictly between 0.2 and 0.8 is moved to 0.2 or 0.8 on the same side, so the model always commits to an answer. Readout temperatures per type are in temps.json (choice 1.1, noul 1.2, score 1.2).

Accuracy by precision

This repository is the 4-bit column. On every benchmark, 8-bit to 5-bit stay within 2 points of 16-bit accuracy and 4-bit within 3 points. 3-bit loses 7 to 21 points and is provided for experimentation.

Accuracy (%)

Benchmark	16-bit	8-bit	6-bit	5-bit	4-bit	3-bit
KoBEST	89.7	89.7	89.6	89.5	87.4	79.7
KorMedMCQA	69.4	69.1	69.6	68.5	66.4	50.8
HAE-RAE	58.3	58.3	56.5	57.5	56.3	40.2
Belebele	89.0	88.8	89.0	88.7	87.8	80.8
KMMLU	56.3	56.1	55.8	56.8	55.3	42.3
CSAT-QA	68.8	68.3	67.7	69.4	66.8	48.3
CLIcK	66.5	66.2	66.9	65.9	65.8	58.8
KMMLU-1035	50.1	49.9	50.5	50.9	50.1	41.5

Headline

Benchmark	16-bit	8-bit	6-bit	5-bit	4-bit	3-bit
KoBEST	92.0	91.9	92.4	91.4	90.8	80.6
KorMedMCQA	78.7	78.7	79.0	77.8	75.5	65.9
HAE-RAE	70.2	69.9	68.9	68.2	68.1	56.9
Belebele	90.3	90.0	89.6	89.2	89.6	82.8
KMMLU	68.5	68.7	68.1	69.7	69.1	55.8
CSAT-QA	76.4	76.4	74.7	74.5	74.5	61.8
CLIcK	72.5	72.2	72.6	73.1	74.5	68.0
KMMLU-1035	60.1	59.9	60.8	61.3	61.1	53.4

Comparisons with other models are in the Xevn-4B evaluation.

Speed

Latency of a single request over HTTP, bf16, Apple M4 Max, 3,000 requests. MLX runtime, one question per request.

Input length	p50	p95
up to 64 tokens	56 ms	64 ms
65 to 128 tokens	88 ms	103 ms
129 to 256 tokens	156 ms	188 ms

Model files

Format	Precision	Size	Speed (p50)	Repository
Transformers	16-bit	9.3 GB		Xevn-4B
MLX	16-bit	8.4 GB	88 ms	Xevn-4B-MLX-bf16
MLX	8-bit	4.5 GB	83 ms	Xevn-4B-MLX-8bit
MLX	6-bit	3.4 GB	82 ms	Xevn-4B-MLX-6bit
MLX	5-bit	2.9 GB	80 ms	Xevn-4B-MLX-5bit
MLX	4-bit	2.4 GB	78 ms	Xevn-4B-MLX-4bit
MLX	3-bit	1.8 GB	77 ms	Xevn-4B-MLX-3bit
GGUF	BF16	8.67 GB		Xevn-4B-GGUF
GGUF	Q8_0	4.61 GB		Xevn-4B-GGUF
GGUF	Q6_K	3.56 GB		Xevn-4B-GGUF
GGUF	Q5_K_M	3.16 GB		Xevn-4B-GGUF
GGUF	Q4_K_M	2.78 GB		Xevn-4B-GGUF

Speed is the MLX p50 over an equal mix of the three input lengths in the Speed table, measured in one session on an Apple M4 Max. On every benchmark, MLX 8-bit to 5-bit stay within 2 points of 16-bit accuracy and 4-bit within 3 points. 3-bit loses 7 to 21 points and is provided for experimentation. The Transformers weights keep the original multimodal layout (9.3 GB); the helper uses the text model only.

Intended use and limitations

Xevn is built for decisions: classification, multiple choice, yes/no judgments, and rating. It is not a chat or text-generation model. It is tuned for Korean; English input works but is not the focus.

Xevn returns probabilities over the options you give it. It does not generate text and cannot answer open questions.
The bundled helpers run one forward pass per question and accept 2 to 26 options and 2 to 10 score levels.
Speed figures are for MLX on an Apple M4 Max with inputs up to 256 tokens.

Citation

@misc{xevn2026,
  title  = {Xevn-4B: a Korean decision model},
  author = {Nhahan},
  year   = {2026},
  url    = {https://huggingface.co/quevirus/Xevn-4B}
}