# OnPrem Branch: Development and custom-AI workstation

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

A coding assistant for your developers and the place to train your own LLM. Your source code and data never leave the company.

Branches are where new work grows, on trees and in git. The number is the graphics memory in GB.

## What it does for your team

- **Completion in the editor:** Suggestions as you type in VS Code or JetBrains, through an OpenAI-compatible endpoint.
- **Ask the codebase:** Explain unfamiliar code, find where things happen, write the missing docs.
- **Coding agents:** Multi-file changes, refactors and tests, run by an agent you can review.
- **Code review:** A second pair of eyes on every pull request, before a human looks.
- **Your code stays yours:** Source code, keys and customer data never reach an outside AI service.
- **Commercial-use models:** Qwen3-Coder models under Apache-2.0, so what they write is yours to ship.
- **Train your own LLM:** Fine-tune an open model on your documents, terminology, tone and rules: a 32B model on Branch 32, 70B to 120B on Branch 96.

## Pick by how you code

Branch 32 for fast completion, chat and mid-size custom models; Branch 96 for coding agents on large codebases and 70B to 120B custom models. The number is the graphics memory in GB.

### OnPrem Branch 32

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: about 30 tokens/s (estimate)

### OnPrem Branch 96

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: about 56 tokens/s (estimate)
- Also runs gpt-oss-120b

## What we install and set up

- An OpenAI-compatible API for your editors and tools
- Qwen3-Coder models sized for the machine
- A private chat app with a login for each developer
- Editor plug-ins set up on your team’s machines
- Fine-tuning tools set up for your own data, with a first trained model
- On-site setup, a training session and a 3-year warranty

## Frequently asked questions

### Does it work with our editors?

Yes. It serves an OpenAI-compatible API, which VS Code and JetBrains coding plug-ins can connect to. We set them up for your team.

### Is our code used to train anything?

No. Everything runs on the machine in your office; nothing is sent out.

### Which tier do we need?

Branch 32 for up to five developers doing completion and chat; Branch 96 for up to twenty, or for coding agents on large codebases.

### Can we train a model on our own data?

Yes. Fine-tuning a 32B model needs about 26 GB and 70B to 120B needs 41 to 65 GB (Unsloth’s published minimums), so Branch 32 and Branch 96 cover both. We set up the tools and deliver a first model with you.

### Can we use the generated code commercially?

The Qwen3-Coder models we install are Apache-2.0 licensed, which allows commercial use.

Request a quote: https://onpremworld.com/en/quote · Not sure? Let Sprout size it for you: https://onpremworld.com/en/find?s=code,customLlm