BuildEng V8 3B — OpenVINO INT8

BuildEng V8 3B OpenVINO INT8 is the optimized INT8 OpenVINO 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 an INT8 OpenVINO conversion of:

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	OpenVINO
Release precision	INT8
BuildEng source model	`Irfanuruchi/qwen2.5-3b-buildeng`
Upstream model	`Qwen/Qwen2.5-3B-Instruct`
Domain	Civil and structural engineering
Primary use	Efficient local engineering-assistant inference on OpenVINO

This release is intended to provide a practical balance between model quality, memory use, and inference performance while preserving the engineering specialization of the full BuildEng V8 3B model.

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

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

Why This Version?

The OpenVINO INT8 release is intended for users who want a practical OpenVINO deployment of BuildEng V8 3B with reduced memory requirements compared with the FP16 release.

Key characteristics include:

INT8 model compression
lower memory requirements than the FP16 OpenVINO release
efficient local inference through OpenVINO
preservation of the BuildEng V8 3B engineering specialization
practical deployment on compatible Intel systems

A 3B INT4 OpenVINO export was also evaluated during development but was not published because its generated output was not sufficiently stable for release. The INT8 release was retained as the preferred compressed OpenVINO version.

Installation

pip install "optimum-intel[openvino]" openvino transformers

Usage

from optimum.intel.openvino import OVModelForCausalLM
from transformers import AutoTokenizer

model_id = "Irfanuruchi/qwen2.5-3b-buildeng-openvino-int8"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = OVModelForCausalLM.from_pretrained(
    model_id,
    device="CPU",
    compile=True,
)

messages = [
    {
        "role": "system",
        "content": (
            "You are BuildEng. Give concise, inspection-first "
            "building engineering guidance."
        ),
    },
    {
        "role": "user",
        "content": (
            "A homeowner reports diagonal cracks near a window corner. "
            "What should be inspected first?"
        ),
    },
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(prompt, return_tensors="pt")

output = model.generate(
    **inputs,
    max_new_tokens=160,
    do_sample=False,
)

new_tokens = output[0][inputs["input_ids"].shape[-1]:]

print(tokenizer.decode(new_tokens, skip_special_tokens=True))

Hardware and Runtime

The example above uses:

device="CPU"

OpenVINO can target different Intel acceleration devices depending on the installed OpenVINO version, hardware, drivers, model representation, and device compatibility.

For this repository:

CPU — primary documented deployment target
Intel GPU — may be used where the model and OpenVINO configuration are compatible
Intel NPU — not specifically claimed or validated for this INT8 release

Users targeting an NPU should follow the OpenVINO requirements for generative-model deployment on their specific hardware and OpenVINO version.

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.

This section is provided as a practical licensing summary. 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 an INT8 OpenVINO 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-openvino-int8

The first stage represents the upstream Qwen model. The second stage represents the BuildEng V8 3B domain fine-tune. The final stage represents the INT8 OpenVINO conversion distributed in this repository.

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

The OpenVINO INT8 files in this repository are modified and converted derivatives of the BuildEng V8 3B release.

Upstream Qwen intellectual-property rights remain subject to the Qwen Research License Agreement.

Author

Irfan Uruçi BuildEng Project

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for Irfanuruchi/qwen2.5-3b-buildeng-openvino-int8

Base model

Qwen/Qwen2.5-3B

Finetuned

Qwen/Qwen2.5-3B-Instruct