Monash University: 3-Node Cluster, DeepSeek V3.1 671B, Cloud Proxies
NuClide Research · 2026-05-01, updated 2026-05-03
Summary
Monash University (Melbourne, Australia) exposes three Ollama nodes on the 118.138.0.0/16 ERC subnet. The primary node (vm-118-138-233-225.erc.monash.edu.au) carries a full DeepSeek V3.1 671B (376.7GB), largest local model in the sweep, alongside two cloud proxies and a qwen3-coder-next:latest (48.2GB). Two secondary nodes run smaller research stacks with active system prompts.
Infrastructure
| Node | IP | Hostname | Version | Models | Cloud |
|---|---|---|---|---|---|
| Primary | 118.138.233.225 | vm-118-138-233-225.erc.monash.edu.au | 0.20.2 | 8 | 2 |
| Node 2 | 118.138.243.239 | vm-118-138-243-239.erc.monash.edu.au | 0.18.3 | 3 | 0 |
| Node 3 | 118.138.243.34 | vm-118-138-243-34.erc.monash.edu.au | 0.19.0 | 3 | 0 |
All on Monash ERC (Education and Research Cluster) subnet.
Model Inventory: Primary Node (118.138.233.225)
| Model | Size | Notes |
|---|---|---|
deepseek-v3.1:latest | 376.7 GB | 671.0B params, largest local model in sweep |
qwen3-coder-next:latest | 48.2 GB | , |
nemotron-cascade-2:latest | 24.3 GB | NVIDIA Nemotron Cascade 2 |
gpt-oss-safeguard:latest | 13.8 GB | gpt-oss 20.9B, safeguard variant, system prompt empty |
kimi-k2.5:cloud | 0 GB | ☁️ Cloud proxy |
minimax-m2.7:cloud | 0 GB | ☁️ Cloud proxy |
gemma4:latest | 9.6 GB | , |
qwen3.5:latest | 6.6 GB | , |
Note: On probe, inference against deepseek-v3.1:latest returns {"error":"model requires more system memory (373.8 GiB) than is available (18.9 GiB)"}. Model is stored on disk but the current primary node lacks sufficient RAM to load it. The model file itself (376.7GB) represents the storage investment; inference would require the full cluster allocation.
Model Inventory: Nodes 2 + 3 (118.138.243.239, 118.138.243.34)
Both nodes carry identical stacks:
| Model | Notes |
|---|---|
deepseek-r1:latest | DeepSeek R1 reasoning |
qwen2.5:latest | Qwen 2.5, default Alibaba system prompt active |
llama3:latest | Meta Llama 3 |
System prompt on qwen2.5: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.", default unmodified installation.
Findings
F1: DeepSeek V3.1 671B Stored (HIGH)
deepseek-v3.1:latest (376.7GB, 671B params) is present on the primary node. Live inference is currently blocked by insufficient system RAM (18.9GB available vs 373.8GB required). However: (a) the model is accessible for download via API, (b) cloud proxies on the same node are fully drained, (c) future memory expansion would make inference live instantly.
F2: Cloud Proxy Portfolio (HIGH)
kimi-k2.5:cloud and minimax-m2.7:cloud on primary node. Both 401 without credential leak.
F3: gpt-oss-safeguard Unconfigured (MEDIUM)
gpt-oss-safeguard:latest (13.8GB) carries the name of a safety-filtered variant, but the system prompt is empty, the filtering was never configured.
F4: 3-Node Cluster, All Injectable (HIGH)
CVE-2025-63389 applies to all 14 models across all three nodes.
Remediation
OLLAMA_HOST=127.0.0.1:11434
systemctl restart ollamaDisclosure
- Discovered: 2026-05-01
- Status: Pending outreach to Monash IT Security (monash.edu.au)