Irfanuruchi/qwen2.5-3b-buildeng-mlx-4bit

BuildEng V8 3B — MLX 4-bit

BuildEng V8 3B MLX 4-bit is an Apple Silicon optimized release of BuildEng V8 3B, a domain-specialized language model focused on civil engineering, structural reasoning, construction workflows, and conservative engineering-assistant behavior.

This release is converted and quantized from:

Irfanuruchi/qwen2.5-3b-buildeng

The BuildEng V8 3B source model is fine-tuned from:

Qwen/Qwen2.5-3B-Instruct

Improved using Qwen.

Model Information

Property	Value
Model	BuildEng V8 3B
Runtime	MLX / mlx-lm
Precision	4-bit
BuildEng source model	Irfanuruchi/qwen2.5-3b-buildeng
Upstream model	Qwen/Qwen2.5-3B-Instruct
Platform	Apple Silicon
Domain	Civil and structural engineering
Primary use	Memory-efficient local engineering-assistant inference on Apple Silicon

The 4-bit MLX release is intended to provide a strong balance between model quality, memory use, model size, and local inference performance on Apple Silicon systems.

Focus Areas

BuildEng is designed around civil, structural, and construction-engineering workflows including:

• building defects and pathology
• reinforced concrete and masonry
• foundations and settlement
• retaining walls
• slabs, beams, and columns
• structural diagnostics
• renovation and existing-structure uncertainty
• structural inspection guidance
• temporary works
• construction sequencing
• construction best practices
• inspection-first reasoning
• conservative engineering recommendations
• educational support for civil and building engineering topics

BuildEng is designed to function as an engineering assistant and to avoid presenting uncertain conclusions as established facts when important project information is missing.

Why This Version?

The MLX 4-bit release is intended for users who want an efficient BuildEng deployment on Apple Silicon with lower memory requirements than the 8-bit MLX release.

Key characteristics include:

• lower memory requirements than the MLX 8-bit release
• smaller model size
• efficient local inference through the MLX framework
• native deployment on Apple Silicon
• a practical balance between model quality and memory use
• a good option for systems with more limited unified memory

For users who prioritize higher model precision over memory efficiency, the MLX 8-bit release is also available.

Installation

pip install -U mlx-lm

Usage

from mlx_lm import load, generate

model, tokenizer = load(
    "Irfanuruchi/qwen2.5-3b-buildeng-mlx-4bit"
)

prompt = """
A homeowner reports diagonal cracks near a window corner.
What should be inspected first?
"""

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=256,
)

print(response)

Example

Prompt

Explain the first inspection steps for settlement cracks in a residential masonry wall.

Example response

Begin by documenting the crack pattern, width, orientation, and location. Check for evidence of ongoing movement, drainage or moisture problems, differential settlement, nearby excavation or loading changes, and whether the affected wall is load-bearing. Repair decisions should follow investigation of the underlying cause rather than treating the crack in isolation.

Hardware

This release is intended for Apple Silicon Macs using the MLX ecosystem.

A system with 8 GB unified memory may be sufficient for lighter use, while 16 GB or more is recommended for more comfortable operation, longer contexts, and running other applications alongside the model.

Suitable systems include:

• MacBook Air with Apple Silicon
• MacBook Pro with Apple Silicon
• Mac mini with Apple Silicon
• Mac Studio
• iMac with Apple Silicon
• Mac Pro with Apple Silicon

Actual memory use and inference performance depend on context length, generation settings, MLX version, and other applications running on the system.

BuildEng Model Family

Release	Repository
Flagship 3B	Irfanuruchi/qwen2.5-3b-buildeng
GGUF Q4_K_M	Irfanuruchi/qwen2.5-3b-buildeng-GGUF-Q4_K_M
GGUF Q8_0	Irfanuruchi/qwen2.5-3b-buildeng-GGUF-Q8_0
GGUF F16	Irfanuruchi/qwen2.5-3b-buildeng-GGUF-F16
MLX 4-bit	Irfanuruchi/qwen2.5-3b-buildeng-mlx-4bit
MLX 8-bit	Irfanuruchi/qwen2.5-3b-buildeng-mlx-8bit
OpenVINO INT8	Irfanuruchi/qwen2.5-3b-buildeng-openvino-int8
OpenVINO FP16	Irfanuruchi/qwen2.5-3b-buildeng-openvino-fp16

Training Dataset

Irfanuruchi/buildeng-v8-3b

License and Commercial Use

BuildEng V8 3B is derived from Qwen/Qwen2.5-3B-Instruct.

The upstream Qwen2.5-3B-Instruct model is distributed under the Qwen Research License Agreement, and the applicable upstream licensing terms continue to apply to this derivative release.

Under the Qwen Research License Agreement, the default license grant covering use, reproduction, distribution, modification, and creation of derivative works is limited to non-commercial research and evaluation purposes.

Commercial use of the upstream Qwen materials requires requesting the appropriate license from Alibaba Cloud.

The Apache-2.0 licenses applicable to some other BuildEng model families therefore must not be interpreted as granting Apache-2.0 rights over the upstream Qwen2.5-3B materials incorporated into this release.

A copy of the Qwen Research License Agreement is included in this repository as:

LICENSE

The required Qwen attribution and information concerning this modified BuildEng derivative are included in:

NOTICE

Qwen Attribution

Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) Alibaba Cloud. All Rights Reserved.

Improved using Qwen.

Commercial BuildEng Alternatives

For deployments that require a BuildEng model derived from an upstream Apache-2.0-licensed model, separate BuildEng variants based on Qwen2.5-1.5B and Qwen2.5-32B are available.

Users should review the applicable license files and ensure that their intended use and redistribution comply with those terms.

Engineering Use Notice

BuildEng is an engineering-assistant model and is not a substitute for professional engineering judgment.

Model output should not be treated as:

• final structural approval
• construction authorization
• engineering certification or sign-off
• a substitute for calculations required by applicable codes or standards
• a substitute for inspection of actual site conditions
• a replacement for review by a qualified or licensed engineer where required

Engineering decisions should be independently verified using appropriate calculations, standards, drawings, site information, project-specific constraints, and professional review.

Modifications and Model Lineage

This repository contains a 4-bit MLX derivative of the fine-tuned BuildEng V8 3B model.

Model lineage:

Qwen/Qwen2.5-3B-Instruct ↓ Irfanuruchi/qwen2.5-3b-buildeng ↓ Irfanuruchi/qwen2.5-3b-buildeng-mlx-4bit

The first stage represents the upstream Qwen model. The second stage represents the BuildEng V8 3B domain fine-tune. The final stage represents the MLX 4-bit conversion distributed in this repository.

BuildEng-specific fine-tuning, engineering specialization, dataset development, evaluation, documentation, and related project contributions were produced by Irfan Uruçi.

Author

Irfan Uruçi BuildEng Project