LDLC WS-TR AI PRO: Local AI Agents with 32GB GPU

Running AI agents without the cloud, on data that never leaves the computer, is the promise of an LDLC workstation designed for heavy models. Whether for confidential transcription, interview preparation, or multi-step automation, the machine leverages 32 GB of VRAM and a 32-core Threadripper, with cost simulation options provided by AMD.

Sensitive processing kept local and subscription costs avoided

Local processing keeps files on the computer rather than on remote servers. This approach aims to avoid the systematic sending of documents to online services and the reliance on subscriptions whose costs can grow with usage. It targets professionals who use AI daily on sensitive data and wish to remain local. The stated goal includes avoiding subscriptions deemed overpriced and the ability to build an internal AI infrastructure tailored to needs. AMD indicates it offers a Tokenomics tool to simulate potential savings based on team size or usage. Additionally, the machine's capacity can be shared among multiple collaborators, depending on the organization set up.

Two agents configured via Hermes for concrete tasks

Two agents have been established with the help of experts via Hermes, relying on the Qwen 3.5 model with 35 billion parameters. One use case involves locally transcribing interviews, with the processor handling the transcription while the graphical resources remain available for other tasks. The agent, named Gus, can then organize the text, extract themes, and retrieve quotes along with their timestamps. A proofreading step remains necessary, especially for proper names. The agent can also analyze several confidential files directly on the workstation. The 32 GB of VRAM helps execute models that accept more context when cross-referencing documents.

A work partner for training, correcting, and orchestrating actions

A second agent has been designed to challenge the user. It assists in preparing interviews in English, makes corrections, and can convert notes in French into a first email in English. Interactions are tailored to the user's level and can be renewed as many times as desired, with no message limit mentioned. In addition to dialogue, the agent handles instructions that include multiple tasks, such as successively performing a transcription, generating a summary from a file, drafting an email, and then organizing the results into dedicated spaces.

Extended use cases: larger models, massive documents, automation

The workstation is presented as capable of handling heavier models and analyzing large volumes of documents. It can contribute to development and coding, automate repetitive operations, and run multiple agents as well as several processes in parallel. An organization can adapt these capabilities into agents dedicated to data analysis, developer assistance, document processing, or administrative task automation.

Hardware configuration: 32-core Threadripper, 32 GB GPU, 64 GB RAM

The LDLC PC WS-TR AI PRO relies on an AMD Ryzen Threadripper 9970X processor with 32 cores, an AMD Radeon AI PRO R9700 graphics card with 32 GB of VRAM, and 64 GB of RAM. This setup aims to execute significant local models while maintaining responsiveness for other tasks. The 32 GB of VRAM facilitates loading large models and providing context, while the processor supports parallel processing. Models are downloaded locally, and agents interact directly with the files present on the machine. According to the presentation, the 32 GB of VRAM and 64 GB of RAM provide headroom to multiply tasks and utilize larger models. This technical proposition positions itself against current AI usage, often reliant on online services generating subscriptions, by highlighting the execution of AI agents locally for professionals handling sensitive content.