Mule Agent is not "a single Java service" — it's a full stack. Beyond the Agent core, it includes a database, knowledge graph, vector search, object storage, document parsing, and more. These are what actually consume most of your RAM and disk.
Below are reference configurations by team size. Key principles:
| Item | Reference config |
|---|---|
| Server | 8-core 32G minimum (Agent + database + knowledge graph + vector search + object storage + text embedding all on one machine. 32G is the floor; go 64G if you have many users or documents) |
| Disk | 512G minimum (knowledge base, chat history, vector index, and media files all accumulate locally) |
| GPU | Not needed (text embedding runs on CPU; the model uses a cloud API) |
| Model | Cloud LLM API (DeepSeek, etc.), pay per use |
| Best for | Small team trials, cross-border e-commerce support, getting started fast |
| Item | Reference config |
|---|---|
| Server | 16-core 64G minimum, stable under concurrent load; consider separating the database and vector search or adding cache |
| Disk | 512G minimum, knowledge base + vector index + media grow quickly |
| GPU | Not needed by default (cloud API); add a GPU host for SGLang if data is sensitive |
| Model | Cloud API primarily; add a GPU to run open-source models if compliance requires |
| Best for | Most office and R&D teams; private deployment for finance/healthcare when compliance demands it |
This size is the dividing line: no strict compliance, cloud API is cheapest; with hard compliance requirements, add a GPU host and go private.
| Item | Reference config |
|---|---|
| Agent server | 8-core 32G (dedicated to the Agent core, no other services mixed in) |
| Database / vector store | Database + vector search + knowledge graph on one separate 16-core 64G machine |
| Object storage | Separate 4-core 8G + large disk (when media/documents are heavy) |
| Disk | 1T minimum total, grows with knowledge base and media |
| GPU | Not needed on cloud; if private, size the GPU host by model |
| Best for | Multi-subsidiary groups, chain retail, cross-border trade — many users, many departments |
At high volume: if you make many calls per day, private SGLang's marginal cost (power + depreciation) can beat pay-per-token API — worth running the numbers at this scale.
| Item | Reference config |
|---|---|
| Agent cluster | Horizontal scale across nodes, 16-core 64G per node; database deployed separately |
| Database / vector store | Database + vector search + knowledge graph on separate HA nodes |
| Object storage | Distributed object storage cluster (multi-node redundancy) |
| GPU | Private SGLang cluster (single/multi GPU, depending on concurrency and model size) |
| Disk | TB-scale to start, grows with knowledge base and audit trails |
| Best for | State-owned enterprises, finance, healthcare, government — strict compliance + large scale |
This size usually has an IT team; we provide onboarding support for deployment and operations (initial implementation + 1–3 months of co-piloting).
If you go private, GPU memory determines how large a model you can run. The table below assumes 4-bit quantization, ~8K context, and a single user:
| Model size | VRAM needed (ref) | Typical GPU | Best for |
|---|---|---|---|
| 7B~8B (e.g. Qwen 7B) | ~6–8 GB | RTX 4060 Ti 16G / 4070 class | Policy lookup, Q&A — light tasks, good enough |
| 14B class | ~10–16 GB | RTX 4070 Ti S / 4080 class | Document writing, spreadsheet cleanup — everyday office |
| 32B class | ~20–32 GB | RTX 4090 24G (quantized) or dual GPU | Close to mid-tier cloud model quality |
| 70B+ class | 48 GB+ | Dual 4090 / A100 class | Plenty of budget, highest demands |
VRAM figures are references and vary with model version. The table is a general industry reference — before deployment we'll produce a config list based on your actual document volume and the model you want to run.
Two factors that significantly increase VRAM — consider them before buying:
| Factor | Impact (reference) |
|---|---|
| Long documents / context | Processing contracts or bids needs 32K-class context. Small models (7B) add only 1~2G — manageable; 32B-class may add 8G or more — the "runs on one 4090" conclusion no longer holds, and you'll need dual GPUs or more VRAM. |
| Concurrent users | When multiple employees ask questions at once, each takes a context-cache slice, stacking VRAM usage. Serving engines like SGLang optimize batching, but total VRAM remains the hard limit. |
So treat the table as a starting point only: one person occasionally looking up policies — pick by the table; dozens of daily users plus long documents — raise the budget a tier, or send us your real scenario — we'll produce a config list based on your document lengths and concurrent users before deployment. Getting this wrong is the most expensive mistake.
| Cloud LLM API | Private Local Model (SGLang) | |
|---|---|---|
| Initial hardware cost | One 8-core 32G + 512G disk to start | Plus a GPU host (thousands to tens of thousands, depending on model size) |
| Long-term model cost | Pay per token to the API provider, only what you use | Power + hardware depreciation, cheaper at high volume |
| Model quality | Cloud flagship models directly, best quality | Open-source is weaker than cloud flagship; 32B covers most office tasks |
| Data leaves your network? | Query content leaves the network (sessions and knowledge base stay local) | Never |
| Works offline? | No (cloud API) | Yes |
| Best for | Cross-border e-commerce support, general office, teams that want to try first | Finance, healthcare, government, state-owned enterprises, R&D code-sensitive teams |
If unsure, start with a cloud API trial: a single 8-core 32G + 512G server/NAS is enough to get running. Let the team use it for two weeks to see real usage and answer quality. Once it proves out, or if compliance demands it, add a GPU host and switch to private — both routes use the same Mule Agent; switching only changes the model endpoint, and your data and skills are fully preserved. No need to start over.
Tell us your team size, main use cases, and data compliance requirements, and we'll give you a deployment config recommendation (with a full hardware list) — free.
Get a free deployment config →