File size: 2,801 Bytes
1197702 65ccee3 1197702 65ccee3 1197702 65ccee3 a06a166 65ccee3 a06a166 65ccee3 d141fcc 65ccee3 a06a166 65ccee3 a06a166 65ccee3 a06a166 65ccee3 a06a166 d141fcc 65ccee3 d141fcc 65ccee3 a06a166 65ccee3 a06a166 65ccee3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 | ---
title: MainStack
emoji: 🛠️
colorFrom: blue
colorTo: indigo
sdk: static
pinned: false
---
# MainStack
### ServiceNow Agentic Delivery — and the open models behind it.
[mainstack.co.uk](https://www.mainstack.co.uk/) · [LinkedIn](https://linkedin.com/company/mainstack-it)
MainStack is a consultancy specializing in ServiceNow Agentic Delivery. We build the AI that does the drafting work of a delivery engagement — and we release it openly. **marvy-1-14B** is our first public model: a fine-tuned LLM that drafts artifacts across the entire ServiceNow delivery lifecycle, from business analysis through Solution Design Documents, user stories, test cases, and validation. It's a first-draft specialist built for the people who do the work — solution architects, business analysts, technical consultants, project managers, and delivery leads.
## 📊 Measured impact
On a **project- and customer-disjoint** held-out test set, `marvy-1-14B` cuts
perplexity on real delivery artifacts by **32% overall vs. the unmodified base
model** — and by **75–86%** on structured artifacts like systems inventories,
requirements, and stakeholder registers. Same weights, adapter on vs. off; the
difference is the fine-tune. See the charts on the
[model card](https://huggingface.co/MainStack/marvy-1-14B#evaluation).
## 🔒 Built responsibly
marvy was trained on real engagement artifacts, which means privacy was a design constraint, not an afterthought. Every training artifact was rigorously anonymized and redacted to **zero residual PII**, verified by an automated leakage scanner. Evaluation was run on a **project- and customer-disjoint** held-out split, so the reported perplexity reflects genuine generalization to unseen work — not memorization. We frame marvy honestly: it accelerates first drafts; it does not replace the judgment of a consultant, and it is not a tool-use or agentic fine-tune.
## 📦 Available formats
| Repo | Format | Use case |
|------|--------|----------|
| [MainStack/marvy-1-14B](https://huggingface.co/MainStack/marvy-1-14B) | Merged FP16 | Full-precision inference and serving |
| [MainStack/marvy-1-14B-lora](https://huggingface.co/MainStack/marvy-1-14B-lora) | LoRA adapter | Compose on top of Qwen2.5-14B-Instruct |
| [MainStack/marvy-1-14B-GGUF](https://huggingface.co/MainStack/marvy-1-14B-GGUF) | GGUF quants | Local runs on Apple Silicon, LM Studio, Ollama |
```bash
ollama run hf.co/MainStack/marvy-1-14B-GGUF:Q4_K_M
```
## About
MainStack specializes in ServiceNow Agentic Delivery, building open, fine-tuned models that draft the artifacts of a delivery engagement — and releasing them to the practitioners who use them. Learn more at [mainstack.co.uk](https://www.mainstack.co.uk/).
*Released under Apache-2.0. Built with Qwen.*
|