# Find your local LLM setup

> OnPrem World · https://onpremworld.com/en/find · Last updated: 2026-10-17

Sprout, the OnPrem World mascot, asks which AI solutions you want, how many people use each and where the machines live, then recommends one complete AI PC per team, each the smallest that covers the job. Prices are by quote.

## The solutions

- **Confidential knowledge base:** A private ChatGPT over contracts, SOPs and internal files, with access by department
- **Finance assistant:** Reads invoices and reports, helps check figures, writes summaries and minutes
- **Video and image generation:** Product photos, ad visuals, short videos and copy for marketing
- **On-prem coding:** A coding assistant for your developers; source code never leaves
- **Custom on-prem LLM training:** Your own model, trained on your data, terminology, tone and rules, including how strict it is
- **Custom video generation:** Image and video models trained on your products, characters and brand style

## What it asks

1. AI that runs on your own computers, so your data never leaves. Which solution do you want to bring in?
2. How capable should the assistant be? How much material should it know? How much video? How many images? What should the coding assistant do? How big a model will you train?
3. How many people will use each one? Just me (One person), A small team (2 to 10 people), A department (10 to 50 people), The whole company (50 people or more)
4. where will the machines live? On a desk, In the office, In a server room

## How it picks

- Confidential knowledge base and finance assistant → OnPrem Root, the office AI PC; video and image generation and custom video generation → OnPrem Bloom, the marketing workstation; on-prem coding and custom on-prem LLM training → OnPrem Branch, the development workstation. Large teams get a shared server.
- The assistant is judged by Qwen3-32B / Qwen3.6-35B-A3B / gpt-oss-20b (everyday) and gpt-oss-120b / Qwen3.5-122B-A10B (close to ChatGPT); coding by Qwen3-Coder-30B-A3B-Instruct and Qwen3-Coder-Next (80B-A3B) at a 32K context, all checked with this site's estimator at Q4_K_M.
- Video: now and then 32 GB, every week 48 GB, every day at full quality 96 GB. Training: a 32B model needs about 26 GB, 70B to 120B needs 41 to 65 GB (Unsloth); visual-model training is sized at 48 GB or more.
- People at once follow each machine’s recommended limit.

## The machines

### 1. Quiet AI mini PC

Silent on a desk. A private assistant that reads your files, for one to three people.

- Chip: AMD Ryzen AI Max+ 395
- Memory: 128 GB unified (shared by CPU and GPU)
- Storage: 2 TB NVMe SSD
- Size: Mini PC, about 3 litres
- Software: Ubuntu or Windows 11 Pro, with a private chat app
- People at once: 3
- Qwen3.6-35B-A3B: fits, about 72 tokens/s (estimate)
- gpt-oss-120b: fits, about 45 tokens/s (estimate)
- Prefer NVIDIA (CUDA)? A DGX Spark with 128 GB fits the same spot and makes images about 2.5 to 4 times faster.

### 2. OnPrem Root 24 (Office AI PC)

The office AI PC: a private assistant for finance, HR and admin, for up to five people.

- Graphics: NVIDIA RTX PRO 4000 Blackwell, 24 GB (ECC), single-slot, 145 W
- Processor: Intel Core Ultra 7 265
- Memory: 64 GB DDR5
- Storage: 1 TB NVMe system + 2× 4 TB mirrored for documents
- Chassis: HP Z2 G1i tower
- Warranty: 3 years, Windows 11 Pro or Ubuntu
- People at once: 5
- Qwen3.6-35B-A3B: fits, about 170 tokens/s (estimate)
- gpt-oss-120b: doesn't fit

### 3. OnPrem Root 32 (Office AI PC)

The office AI PC for up to ten people, with room for longer documents.

- Graphics: NVIDIA RTX PRO 4500 Blackwell, 32 GB (ECC), 200 W
- Processor: Intel Core Ultra 9 285
- Memory: 128 GB DDR5
- Storage: 1 TB NVMe system + 2× 8 TB mirrored for documents
- Chassis: HP Z2 G1i tower
- Warranty: 3 years, Windows 11 Pro or Ubuntu
- People at once: 10
- Qwen3-32B: fits, about 30 tokens/s (estimate)
- gpt-oss-120b: doesn't fit

### 4. OnPrem Bloom 32 (Marketing AI workstation)

The marketing AI workstation for product images, ad visuals and the occasional short video.

- Graphics: NVIDIA RTX PRO 4500 Blackwell, 32 GB (ECC), 200 W
- Processor: Intel Core Ultra 9 285
- Memory: 128 GB DDR5
- Storage: 2 TB NVMe system + 8 TB for projects
- Chassis: HP Z2 G1i tower
- Warranty: 3 years, Windows 11 Pro
- People at once: 10
- Qwen3-32B: fits, about 30 tokens/s (estimate)
- gpt-oss-120b: doesn't fit

### 5. OnPrem Bloom 48 (Marketing AI workstation)

The marketing AI workstation for video every week: images, short films and copy, made in-house.

- Graphics: NVIDIA RTX PRO 5000 Blackwell, 48 GB (ECC), 300 W
- Processor: Intel Core Ultra 9 285
- Memory: 128 GB DDR5
- Storage: 2 TB NVMe system + 8 TB for projects
- Chassis: HP Z2 G1i tower
- Warranty: 3 years, Windows 11 Pro
- People at once: 10
- Qwen3-32B: fits, about 43 tokens/s (estimate)
- gpt-oss-120b: doesn't fit

### 6. OnPrem Bloom 96 (Marketing AI workstation)

The marketing AI workstation for daily, full-quality video, with a ChatGPT-class model loaded beside it.

- Graphics: NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96 GB (ECC), 600 W
- Processor: Intel Core Ultra 9 285
- Memory: 128 GB DDR5
- Storage: 2 TB NVMe system + 16 TB for projects
- Chassis: HP Z2 G1i tower
- Warranty: 3 years, Windows 11 Pro
- People at once: 10
- Qwen3-32B: fits, about 56 tokens/s (estimate)
- gpt-oss-120b: fits, about 224 tokens/s (estimate)

### 7. OnPrem Branch 32 (Development and custom-AI workstation)

Fast code completion and chat for up to five developers, and room to train a mid-size custom model.

- Graphics: NVIDIA RTX PRO 4500 Blackwell, 32 GB (ECC), 200 W
- Processor: Intel Core Ultra 9 285
- Memory: 128 GB DDR5
- Storage: 2 TB NVMe
- Chassis: HP Z2 G1i tower
- Warranty: 3 years, Ubuntu with an OpenAI-compatible API
- People at once: 5
- Qwen3-32B: fits, about 30 tokens/s (estimate)
- gpt-oss-120b: doesn't fit

### 8. OnPrem Branch 96 (Development and custom-AI workstation)

Coding agents on large codebases for up to twenty developers, and room to train 70B to 120B custom models.

- Graphics: NVIDIA RTX PRO 6000 Blackwell Max-Q, 96 GB (ECC), 300 W
- Processor: Intel Core Ultra 9 285
- Memory: 128 GB DDR5
- Storage: 4 TB NVMe
- Chassis: HP Z2 G1i tower
- Warranty: 3 years, Ubuntu with an OpenAI-compatible API
- People at once: 20
- Qwen3-32B: fits, about 56 tokens/s (estimate)
- gpt-oss-120b: fits, about 224 tokens/s (estimate)

### 9. Team AI workstation

A ChatGPT-class assistant for a whole team, and video at full quality.

- Graphics: NVIDIA RTX PRO 6000 Blackwell Max-Q, 96 GB (room for a second)
- Processor: Intel Xeon W or AMD Threadripper PRO
- Memory: 256 GB ECC
- Storage: 2 TB NVMe system + 8 TB data
- Chassis: Workstation tower, 1600 W
- Warranty: 3 years, Ubuntu with team accounts
- People at once: 50
- Qwen3-32B: fits, about 56 tokens/s (estimate)
- gpt-oss-120b: fits, about 224 tokens/s (estimate)

### 10. Department AI server

Hundreds of people, the biggest open models, and room to fine-tune your own.

- Graphics: 4× NVIDIA RTX PRO 6000 Blackwell Max-Q, 384 GB in total
- Processor: AMD Threadripper PRO, 32 cores or more
- Memory: 512 GB ECC
- Storage: 2 TB NVMe system + 16 TB data
- Chassis: Rack or tower, dual 1600 W
- Warranty: 3 years on site, Ubuntu with accounts and monitoring
- People at once: 200
- Qwen3-32B: fits, about 56 tokens/s (estimate)
- gpt-oss-120b: fits, about 224 tokens/s (estimate)
- Or an NVIDIA DGX Station GB300 (252 GB on one chip), by project quote.

Start: https://onpremworld.com/en/find · Request a quote: https://onpremworld.com/en/quote