GH Apex (4B)
GH Apex is part of GH AI, the open model family from Growthackers, an Italian growth-marketing agency: GH Spark (0.6B), GH Pulse (1.7B) and GH Apex (4B). Best for: strategy and long structured answers.
It is a LoRA fine-tune of Qwen/Qwen3-4B (Apache-2.0) on Growthackers' public growth-marketing knowledge
(playbooks, metrics library, glossary, acquisition / content / SEO / CRO / AI-tool notes) plus an open
general-chat slice (HuggingFaceTB/smoltalk, Apache-2.0). Trained on a single laptop CPU.
Evaluation (base → GH Apex)
| Metric | Qwen3-4B | GH Apex |
|---|---|---|
| Loss on held-out GH notes (lower is better) | 2.908 | 2.047 |
| ARC-Easy (zero-shot) | 77.0% | 83.0% |
| ARC-Challenge (zero-shot) | 52.0% | 57.0% |
Usage
ollama create gh-apex -f Modelfile && ollama run gh-apex
llama-server -m gh-apex-Q4_K_M.gguf --jinja -c 8192 # OpenAI-compatible API
Use non-thinking mode (enable_thinking=False); the GH system prompt is baked into the Modelfile.
Training data and privacy
Only generic-knowledge notes were used. Notes mentioning any client were dropped; people's names, emails, phone numbers, IBANs, ad-account IDs and API keys were scrubbed; HR, finance, partner, legal and client-success material was excluded entirely.
Limitations
Small model: it can invent facts, statistics and URLs, and its arithmetic is unreliable. No internet
access. Verify figures before using its output with clients.
Known issue: asked for contact details it can still guess Growthackers addresses (e.g. a
growthackers.it site/email); the official site is https://gh.company.
- Downloads last month
- 22