PC for local AI: 2026 configurations

Why run AI locally?

In 2026, it is no longer necessary to use a cloud service to use a high-performance language model. Open-source models (Llama, Mistral, Qwen, gpt-oss, DeepSeek...) run directly on a machine installed in your premises, without any data going to an external server.

For professions subject to professional secrecy or sensitive data (lawyers, notaries, accountants, doctors, HR, consulting firms), this is what makes AI usable on a daily basis: your files stay with you.

• Confidentiality: documents and queries stay on your local network.
• No per-user subscription: once the machine is in place, the cost per query is almost zero.
• Offline operation: the AI remains available even without an internet connection.
• Complete control: choice of models, RAG on your documents, agents, fine-tuning.

Key criteria for an AI PC configuration

1. The memory available for the model

VRAM (the graphics card memory) determines the size of the models you can load. A model must fit entirely in VRAM to run at full speed: if it spills over into system memory, generation becomes much slower.

A simple rule to estimate the need: in 4-bit quantization (a compression of the model weights that preserves most of its quality), allow about 0.6 GB per billion parameters, plus a few GB for context. In 8-bit, double it. In full precision (FP16), allow about 2 GB per billion parameters.

Indicative values for a common 4-bit quantization (type Q4_K_M) and a moderate context. A long context or several simultaneous users add several GB.

MoE (Mixture of Experts) models like Llama 4 or Qwen3 235B occupy a lot of memory, but only activate a part of their parameters for each generated word. They are therefore faster than a classic model of the same size.

2. Memory bandwidth

Once the model is loaded, the generation speed depends mainly on the speed at which the memory is read. An RTX 5090 or an RTX PRO 6000 exceeds 1.7 TB/s, while the unified memory of the GB10 runs around 273 GB/s. Consequence: on a model that fits in a graphics card, a workstation with a dedicated GPU responds faster. The GB10 takes the lead as soon as the model exceeds the VRAM of a single card.

3. System memory, processor and storage

System memory counts for RAG pipelines, processing large documents and working with multiple people. 32 GB is a serious minimum, 64 GB and more becomes useful as soon as several tasks run in parallel. The processor counts less than the GPU, except in multi-GPU: you need a platform with enough PCIe lanes, like the Threadripper PRO. Finally, plan for at least a 1 TB NVMe SSD: models quickly weigh several dozen GB each.

Radiance Systems AI machine comparison

Base configuration price including VAT, automatically updated from our store. All machines remain configurable. The details of each model, with the uses it corresponds to, are found on the page of our range of local AI servers and workstations.

CoreIA Radeon AI PRO R9700: the entry-level professional AI workstation

The CoreIA Radeon AI PRO R9700 is the right entry point to use local AI seriously on a daily basis, on a workstation that is also used for everything else.

• GPU: AMD Radeon AI PRO R9700, 32 GB of VRAM
• Processor: AMD Ryzen 9 9900X, 12 cores / 24 threads
• Memory: 32 GB DDR5, expandable up to 256 GB
• Storage: 1 TB NVMe SSD
• System: Windows 11 Professional

With 32 GB of VRAM, it comfortably runs text models up to 30B in 8-bit (Qwen3 14B, Phi-4), the 8B with a long context, via Ollama, LM Studio or llama.cpp. Being an AMD card, it does not have CUDA (ROCm or Vulkan): image generation with FLUX or Stable Diffusion works less well on it, so a CoreIA RTX 5090 is better. It is the typical machine of a lawyer, an accountant or an independent who wants an assistant on their own documents.

CoreIA RTX 5090: the high-end AI workstation

The CoreIA RTX 5090 relies on the RTX 5090 and its 32 GB of VRAM, the fastest consumer card for local AI. It runs models up to 32B at full speed, with room for context.

• GPU: NVIDIA GeForce RTX 5090, 32 GB of VRAM
• Processor: AMD Ryzen 9 9900X, 12 cores / 24 threads
• Memory: 64 GB DDR5, expandable up to 256 GB
• Storage: 2 TB Gen 5 NVMe SSD, expandable
• System: Windows 11 Professional

It is the reference for AI development, demanding RAG pipelines, light fine-tuning (LoRA, QLoRA) and creation (images, video, 3D). The GPU remains configurable: you can go for a 96 GB Blackwell RTX PRO 6000 if your models require it, or add a second card.

CoreIA Threadripper PRO 4 GPU and up to 4 RTX PRO 6000

The CoreIA Threadripper PRO 4 GPU is our multi-GPU platform. The Threadripper PRO brings the PCIe lanes and memory necessary to power up to 4 RTX PRO 6000 Blackwell of 96 GB, i.e., up to 384 GB of VRAM with 4 cards.

• One card (96 GB): Llama 3.3 70B in 8-bit, gpt-oss-120b, Llama 4 Scout.
• Two cards (192 GB): Qwen3 235B, or several models served in parallel to a team.
• Three cards (288 GB): Llama 4 Maverick (400B), fine-tuning of medium models.
• Four cards (384 GB): the largest models, or several large models served in parallel to a team.

You can start with a single card and add more later: the machine is designed to evolve. It is the choice suited to an SME that wants an AI assistant shared by several dozen people, or to an R&D team. For sizing according to your staff, consult our guide on the local AI server in France.

Rack AI servers

If you have a computer rack, our rack servers integrate into your server room and serve AI to the entire network.

• CoreIA 2x RTX PRO 6000: 192 GB of VRAM to serve large models or several models at once, do fine-tuning or host an enterprise RAG.
• CoreIA 2x RTX 5090: 64 GB of VRAM and very fast generation, for models up to 70B in 4-bit.

With two cards, the model is distributed between the GPUs by the inference software (vLLM, llama.cpp), which adds their memory.

ASUS Ascent GX10 mini-server (NVIDIA GB10)

The ASUS Ascent GX10 is the ASUS version of the NVIDIA DGX Spark, based on the GB10 Grace Blackwell chip. This is not a desktop PC: it is an AI mini-server of 150 × 150 × 51 mm, designed for inference, RAG, agents and development.

• 128 GB of unified memory LPDDR5X, shared between the processor and the GPU
• Up to 1 petaFLOP of AI power in FP4 precision
• Grace ARM processor with 20 cores
• DGX OS (based on Ubuntu) with CUDA, PyTorch and NVIDIA tools pre-installed
• ConnectX-7 Network: two GX10 can be connected to run models of around 235B

Its 128 GB allows loading models that no consumer card can contain, like gpt-oss-120b or Llama 4 Scout. In return, its memory bandwidth is lower than that of an RTX 5090: on a model that fits in 32 GB, a CoreIA RTX 5090 generates faster.

Which configuration to choose according to your use?

Three questions are enough: what model size are you aiming for, how many people will use it, and do you need Windows on the same machine?

Are you hesitating between a Mac and a PC? We compared the two in our Mac Studio vs PC RTX for local AI comparison. And for a focus on language models, see our guide which PC for a local LLM.

Deployment and support

An AI machine is only useful if it is well installed. This is what distinguishes us from a simple hardware reseller:

• AI turnkey pack: models installed, chat interface ready to use and tests performed before delivery.
• On-site installation in France and Europe: networking, access configuration and handover.
• Custom development: n8n automations and Python applications connected to your documents and tools.
• Specialized AI support: the same contact person for hardware and software.

Frequently Asked Questions