Agricultural University of Athens: 142GB Qwen3-235B MoE, Dual-Embedding RAG
NuClide Research · 2026-05-03
Summary
afa4pc19.aua.gr (143.233.187.19), Agricultural University of Athens (Γεωπονικό Πανεπιστήμιο Αθηνών, AUA), runs Ollama v0.18.2 with a 5-model stack anchored by qwen3:235b-a22b-instruct-2507-q4_K_M, the Qwen3 235B MoE model (235.1B total params, 22B active, July 2025 instruction variant) at 142GB. A dual-embedding RAG pipeline (BGE-M3 + nomic-embed-text) is running alongside DeepSeek-R1:32B and Llama3.3:70B.
Infrastructure
| Field | Value |
|---|---|
| IP | 143.233.187.19 |
| Hostname | afa4pc19.aua.gr |
| Organization | Agricultural University of Athens (AUA) |
| Network | AUA Greece (143.233.0.0/16) |
| Country | Greece |
| Ollama version | 0.18.2 |
| Open port | 11434 (public) |
Model Inventory
| Model | Size | Notes |
|---|---|---|
qwen3:235b-a22b-instruct-2507-q4_K_M | 142GB | 235.1B params, 22B active (MoE), July 2025 instruction variant, Q4_K_M |
llama3.3:70b | 42GB | Llama 3.3 70B |
deepseek-r1:32b | 19GB | DeepSeek reasoning model |
bge-m3:latest | 1GB | BGE-M3 multilingual RAG embedding |
nomic-embed-text:latest | , | Nomic embedding, secondary RAG layer |
Total local storage: ~204GB+
Findings
F1: Qwen3 235B: Largest MoE Model on Greek Academic Infrastructure (HIGH)
qwen3:235b-a22b-instruct-2507-q4_K_M is a 235.1B parameter mixture-of-experts model, 22B parameters active per inference pass, running on agricultural university hardware. The 2507 tag indicates the July 2025 instruction fine-tune (the most current Qwen3 variant). At Q4_K_M quantization, the 142GB weight file requires significant GPU memory. The “afa4pc19” hostname suggests this is a department workstation (PC #19, AFA department, possibly Applied Forestry, Agricultural Sciences, or similar).
Unauthenticated inference against a 235B MoE constitutes significant compute theft at scale.
F2: Dual-Embedding RAG Pipeline (HIGH)
Both bge-m3:latest (multilingual, 500M+ params) and nomic-embed-text:latest are present simultaneously, a two-embedding RAG architecture. BGE-M3 supports 100+ languages, suggesting multilingual document retrieval. The combination with DeepSeek-R1 and Qwen3-235B indicates an active research-grade RAG system.
Documents indexed into the vector store (research papers, agricultural datasets, lab reports, student theses) are queryable via the unauthenticated API.
F3: CVE-2025-63389 (CRITICAL)
All 5 models injectable via unauthenticated /api/create. The 235B model’s context handling is overwritable.
Remediation
OLLAMA_HOST=127.0.0.1:11434
systemctl restart ollamaDisclosure
- Discovered: 2026-05-03
- Status: Pending outreach to AUA IT Security (aua.gr) / GRNET (GEANT)