⚠️ PUBLIC PREVIEW — NOT OPEN SOURCE. This is a controlled evaluation release. Commercial use, hosted inference, derivative model creation, fine-tuning, distillation, and redistribution are not permitted without a separate written commercial licence from Rotex Technologies Ltd. See license.md.

---

RotexMech-8B — Public Preview GGUF Model

1. Executive Summary

RotexMech-8B is a public preview GGUF release developed by Rotex Technologies Ltd — a UK-based digital manufacturing, engineering services specialising in CFD, FEA, and technical consultancy across the nuclear, aerospace, automotive, marine, wind, solar, energy storage, and oil & gas sectors.

RotexMech-8B is fine-tuned from the Qwen3-8B base model, with a training and evaluation focus on mechanical engineering domains — structural integrity, materials and corrosion, thermal and fluid systems, dynamics and tribology, manufacturing and DFMA, reliability and lifecycle, rotating equipment diagnostics, and pressure systems integrity.

This release provides a Q4_K_M quantised model (RotexMech-8B-q4_k_m.gguf) intended to demonstrate the capability of our wider commercial model family, whilst preserving our proprietary full‑precision weights, training pipeline, domain‑specific datasets, and deployment optimisations.

✅ This model is suitable for:

• Local, offline evaluation and experimentation
• Technical benchmarking against publicly reported metrics
• Early-stage customer and partner feedback
• Investor and technical due diligence
• Edge and air-gapped inference testing
• Academic and industry technical review

❌ This model is not licensed for:

• Commercial deployment or production use
• Hosted inference, SaaS, or API services
• Fine‑tuning, distillation, or adapter training
• Derivative model creation of any kind
• Redistribution at scale or mirroring
• Use as a component in a competing product or service

For commercial access, evaluation under NDA, or partnership discussions, contact: 📧 sales@rotextechnologies.com

2. Why This Model Exists

Most current models in this parameter class are either:

• too large for practical edge and offline deployment;
• too generic for specialist industrial and engineering use;
• costly to host and difficult to validate in regulated environments;
• trained on data of unknown provenance, creating unacceptable IP and compliance risk;
• unsuitable for controlled, auditable engineering workflows.

Rotex Technologies Ltd is building a model family focused on:

• Domain precision — trained and evaluated on curated mechanical and industrial engineering literature and case material, spanning structural, thermal, materials, dynamics, manufacturing, reliability, and rotating/pressure equipment topics;
• Deployability — designed for edge, air-gapped, and resource-constrained environments from the outset;
• Auditability — built for workflows where traceability, provenance, and controlled outputs matter;
• Data security and privacy — fully air-gapped development pipeline; all processes, from raw document ingestion to model build, are isolated from the internet, ensuring complete confidentiality of datasets and training corpora;
• Commercial clarity — a transparent licensing model that respects both open evaluation and proprietary IP.

This public preview is a release in our production model pipeline — a controlled, quantised snapshot sufficient for meaningful technical evaluation without compromising our proprietary training assets, architecture, or development IP.

3. Target Audience

This preview is intended for:

Audience	Purpose
Technical evaluators	Assess model capability, quality, and behaviour
Engineering teams	Test domain-specific performance on representative mechanical engineering tasks
Enterprise innovation groups	Evaluate fit within controlled internal workflows
Investors & due diligence teams	Conduct independent technical assessment
Research partners	Explore collaboration potential
Early design partners	Inform commercial roadmap and integration requirements

It is not intended for, and not licensed for, production deployment of any kind.

4. Model Details

Field	Detail
Model name	RotexMech-8B
File name	RotexMech-8B-q4_k_m.gguf
Base model	Qwen3-8B
Release type	Public preview (evaluation only)
File format	GGUF
Quantisation	Q4_K_M
Primary language	English
Domain focus	Mechanical engineering — structural integrity, materials, thermal/fluids, dynamics, manufacturing/DFMA, reliability, rotating equipment, pressure systems
Release date	3 August 2026
Pipeline position	Public preview release; higher-precision variants (BF16, Q8_0) available to qualified commercial customers
Licence	Model Public Preview Licence v1.0
Commercial use	Not permitted without written licence from Rotex Technologies Ltd
Developer	Rotex Technologies Ltd, United Kingdom

5. What Is Included — and What Is Not

✅ Public in this release:

• Q4_K_M GGUF model file
• Inference instructions for llama.cpp and compatible runtimes
• Evaluation prompts and example outputs
• Safety, limitation, and known-issue notes
• Public benchmark summary
• Full licence terms and acceptable use policy

❌ Not included in this release:

• Higher-precision model variants (BF16)
• Higher-quality quantisations (Q8_0)
• Training, fine‑tuning, or alignment datasets
• Fine‑tuning recipes, hyperparameters, or training code
• Internal evaluation harness and proprietary benchmarks
• Commercial deployment configuration and system prompts
• Domain‑specific commercial variants
• Any component that would enable derivative model creation

6. Permitted & Prohibited Uses

✅ Permitted (subject to licence terms):

• Local, non‑commercial evaluation and experimentation
• Technical benchmarking and comparison
• Academic and industry review and publication of fair results
• Internal proof‑of‑concept assessment
• Investor and partner due diligence

❌ Prohibited (without separate written commercial licence):

• Production deployment of any kind
• Commercial use, revenue generation, or cost reduction
• SaaS, API, or hosted inference services
• Fine‑tuning, distillation, merging, or adapter creation
• Training or improving any other model (competing or otherwise)
• Redistribution, mirroring, or repackaging at scale
• Any use violating the Acceptable Use Policy

Full terms: license.md

7. Quick Start

Prerequisites

• llama.cpp (built with GPU acceleration recommended)
• Or any GGUF-compatible runtime (Ollama, LM Studio, etc.)

Basic inference with llama.cpp

# Download the model file from this repository
# Then run:

./llama-cli \
  -m RotexMech-8B-q4_k_m.gguf \
  -p "Explain the key technical considerations for fatigue analysis in pressure vessel design." \
  -n 512 \
  --temp 0.6 \
  --ctx-size 2048

Interactive mode

./llama-cli \
  -m RotexMech-8B-q4_k_m.gguf \
  --interactive \
  --temp 0.6 \
  --ctx-size 2048

With Ollama

Create a Modelfile:

FROM ./RotexMech-8B-q4_k_m.gguf

Then:

ollama create rotexmech-8b-preview -f Modelfile
ollama run rotexmech-8b-preview

Suggested evaluation prompts

"Summarise the key failure modes of a welded joint in cooling water pipes under cyclic thermal loading."

"Outline a root-cause investigation approach for bearing overheating in a centrifugal pump."

"Explain the difference between hydrostatic and pneumatic pressure testing for a vessel returning to service after a repair, and the key precautions for each."

8. Evaluation & Benchmarking

We encourage independent benchmarking and technical review. When publishing results:

• ✅ Attribute the model as: "RotexMech-8B by Rotex Technologies Ltd, used under the Model Public Preview Licence v1.0."
• ✅ Be transparent about your evaluation methodology, prompt format, and sampling parameters.
• ✅ Publish both strengths and limitations.
• ❌ Do not imply endorsement, partnership, or certification by Rotex Technologies Ltd.
• ❌ Do not disclose any information that would enable reverse engineering or derivative model creation.
• ❌ Do not use model outputs to train or improve competing models.

8.1 MMLU-Pro benchmark

MMLU-Pro is a public, reasoning-focused extension of MMLU spanning 14 academic and technical domains. The figures below combine self-reported RotexMech-8B results with published third-party figures from qwen.ai and TIGER-AI-Lab, using the "Overall", "Engineering", and "Math" subsets.

Figure 1 — MMLU-Pro benchmark: Math, Engineering, and Overall subset scores.

Model	Overall	Engineering	Math	Data source
Qwen3-8B-Instruct	0.683	0.567	0.676	qwen.ai
Qwen2.5-14B	0.637	0.495	0.679	TIGER-AI-Lab
Qwen3-8B (base)	0.594	0.484	0.602	qwen.ai
RotexMech-8B	0.595	0.500	0.745	Self-reported
Qwen2.5-7B	0.552	0.416	0.523	qwen.ai
Llama-3.1-8B-Instruct	0.443	0.297	0.438	TIGER-AI-Lab

8.2 In-house mechanical engineering benchmark

We also assess models against an in-house mechanical engineering benchmarking tool: 100 mechanical engineering questions spanning eight topic areas (structural/fatigue/fracture, materials/corrosion/joining, thermal/fluids/sealing, dynamics/vibration/tribology, manufacturing/DFMA/quality, reliability/safety/lifecycle, rotating equipment diagnostics, and pressure systems integrity), scored by a judge model out of 100.

Figure 2 — Overall mean judged score (100 in-house mechanical engineering questions). Source: in-house mechanical engineering benchmarking tool.

Model	Mean score / 100	Non-zero coverage	Rank
Qwen2.5-14B	58.2	88%	1
Llama-3.1-8B	47.0	73%	2
RotexMech-8B	43.1	56%	3
Qwen2.5-7B	40.2	59%	4
Qwen3-8B-Instruct	31.5	40%	5
Qwen3-8B (base)	28.3	35%	6

Figure 3 — Capability profile by mechanical engineering topic. Source: in-house mechanical engineering benchmarking tool. Read as a capability shape comparison rather than a standalone ranking.

We welcome feedback on evaluation methodology, benchmark design, and domain coverage at: 📧 info@rotextechnologies.com

9. Known Limitations

As a public preview Q4_K_M quantised model, RotexMech-8B has the following known limitations:

• Quantisation degradation — Q4_K_M quantisation intentionally reduces fidelity compared to our full‑precision commercial models. Expect some degradation on nuanced technical reasoning and numerical precision.
• No production safety guarantees — This model has not undergone production safety alignment, red‑teaming, or adversarial robustness testing at commercial grade.
• Benchmark coverage — On the in-house mechanical engineering benchmark, RotexMech-8B returned a lower non-zero response coverage (56%) than several comparison models, and trails Qwen2.5-14B and Llama-3.1-8B on rotating equipment diagnostics. Results should be read topic-by-topic rather than from the overall score alone.
• Domain coverage — The model's knowledge reflects its training distribution. It may perform less well on topics outside mechanical and industrial engineering domains.
• Hallucination — Like all large language models, RotexMech-8B may generate plausible but factually incorrect output. Never rely on model output for safety‑critical decisions without independent verification.
• No regulatory approval — This model has not been assessed or approved by any regulatory body for use in safety‑critical or regulated applications.
• Language — Optimised for English. Performance in other languages is untested and may be unreliable.

This model is an evaluation artefact. It is not a production‑ready, safety‑assured product.

10. Safety & Responsible Disclosure

We take safety concerns seriously. If you discover:

• Generation of hazardous, dangerous, or illegal content
• Instructions that could cause physical harm
• Bypass of safety guardrails
• Biased, discriminatory, or harmful outputs
• Vulnerabilities that could be exploited
• Copyright or IP concerns in model outputs

Please report responsibly to: 📧 info@rotextechnologies.com

We commit to acknowledging reports within 5 working days and providing a substantive response within 20 working days. We request a 90‑day embargo before public disclosure of vulnerabilities to allow for assessment and mitigation.

Do not use this model to generate, distribute, or facilitate content that would violate applicable laws in the United Kingdom, European Union, United States, People's Republic of China, or any other relevant jurisdiction.

11. Citation

If you reference RotexMech-8B in academic or technical publications:

@software{rotex_rotexmech8b_2026,
  author = {{Rotex Technologies Ltd}},
  title = {RotexMech-8B — Public Preview GGUF Model},
  year = {2026},
  publisher = {Rotex Technologies Ltd},
  url = {https://huggingface.co/rotextechnologies/rotexmech-8B},
  note = {Model Public Preview Licence v1.0. Q4_K_M quantisation. Fine-tuned from Qwen3-8B.}
}

Suggested in-text attribution: "RotexMech-8B by Rotex Technologies Ltd, used under the Model Public Preview Licence v1.0."

12. Commercial Access

This public preview is not licensed for commercial use. Our commercial offering includes:

• Higher-precision (BF16) and higher-quality quantised (Q8_0) model variants
• Domain‑specific fine‑tuned variants
• Commercial deployment licence
• SLAs, support, and integration assistance
• Access to proprietary evaluation tooling
• Custom development and bespoke model adaptation

For commercial enquiries, partnership discussions, or evaluation under NDA: 📧 sales@rotextechnologies.com 🌐 www.rotextechnologies.com

See COMMERCIAL_USE.md for more detail.

13. Repository Documents

Document	Description
license.md	Full legal terms — Model Public Preview Licence v1.0
ACCEPTABLE_USE_POLICY.md	Acceptable use policy and prohibited use categories
COMMERCIAL_USE.md	Commercial licensing pathway and FAQ

---

14. Acknowledgements

RotexMech-8B is fine-tuned from the Qwen3-8B base model. Benchmark comparisons in Section 8.1 draw on published and self-reported data from qwen.ai and TIGER-AI-Lab (MMLU-Pro leaderboard). We gratefully acknowledge these organisations for their open publication of model architectures and benchmark results, which support transparent, independent evaluation across the field. Inclusion of third-party figures does not imply endorsement of RotexMech-8B by qwen.ai or TIGER-AI-Lab.

---

© Rotex Technologies Ltd. All rights reserved. Rotex Technologies Ltd is a company registered in England and Wales.