01
Justified components
Every GPU, VRAM and PSU choice is explained for a real local AI workload.
Reference builds
for AI rigs.
Three educational reference builds — GPU, VRAM, sizing. What a good local AI rig looks like, with no commercial bias.
01 · Sovereignty
Sensitive data (legal, medical, HR, R&D) never leaves your walls. GDPR and AI Act compliant.
02 · Bounded cost
A Pro rig pays back in ~6 months versus €1500/month of API at scale. Marginal cost near zero afterwards.
03 · Updated models
How to keep up with new open-weight models, pick the right quantization and test a launch: our guides walk you through it step by step.
trust
What makes these builds reliable.
Localia is an independent resource: no stock, no fake benchmark. These builds are a reference to assemble the right rig yourself.
02
No pricing shown
Localia does not sell: these are references to assemble yourself, whose cost you check with the reseller of your choice.
03
Documented performance
Which models run and their quantization are specified tier by tier.
04
Benchmarks to come
Real tokens/s measured on physical hardware will be published as soon as possible.
catalog
Three reference builds. Build them yourself.
reference builds · no pricing, Localia does not sell
// 01 · Solo dev · researcher · internal experiments
Starter
First solo AI rig
GPU · VRAM
1× RTX 509032 GB
What fits in this VRAM
- Mistral 7BFP16 · 16 GB
- Qwen 2.5 32BQ5 · 25 GB
- Llama 3.3 70Bdoes not fit
- Mistral Large 123Bdoes not fit
best quantization that fits · calculator engine
Spec
- ·AMD Ryzen 7 7700X (8C / 16T)
- ·32 GB DDR5-6000
- ·1 TB NVMe Gen 4
- ·850 W 80+ Gold PSU
- ·Component manufacturer warranties
Runs for example
- ▸Qwen 3 30B-A3B · Q5
- ▸Qwen 2.5 32B · Q5
- ▸Gemma 4 31B · Q5
- ▸Mistral 7B · FP16
- ▸DeepSeek R1 Distill 32B · Q5
Software
- ·Ubuntu 24.04 LTS
- ·NVIDIA drivers + CUDA
- ·Ollama + Open WebUI (open-source, to install)
- ·Optimised llama.cpp build
// 02 · Agency · SMB · firm · data team
popularPro
The team RAG rig
GPU · VRAM
2× RTX 509064 GB
What fits in this VRAM
- Mistral 7BFP16 · 16 GB
- Qwen 2.5 32BQ8 · 36 GB
- Llama 3.3 70BQ5 · 54 GB
- Mistral Large 123BQ3 · 60 GB · tight
best quantization that fits · calculator engine
Spec
- ·AMD Ryzen 7 7700X (8C / 16T)
- ·64 GB DDR5-4800 ECC UDIMM
- ·1 TB NVMe Gen 4
- ·1500 W 80+ Platinum PSU
- ·Component manufacturer warranties
Runs for example
- ▸Llama 3.3 70B · Q5
- ▸DeepSeek R1 Distill 70B · Q5
- ▸Qwen 2.5 72B · Q5
- ▸Mixtral 8x7B · Q8
- ▸Mistral Large 123B · Q3
Software
- ·Everything in Starter +
- ·vLLM with OpenAI-compatible server
- ·RAG kit: LlamaIndex + Qdrant configured + README
- ·CLI monitoring (nvidia-smi, htop, ollama logs)
// 03 · SME · law firm · medical · mid-market · public
Enterprise
AI sovereignty for SMEs
GPU · VRAM
2× RTX A6000 NVLink96 GB
What fits in this VRAM
- Mistral 7BFP16 · 16 GB
- Qwen 2.5 32BFP16 · 72 GB
- Llama 3.3 70BQ8 · 78 GB
- Mistral Large 123BQ5 · 95 GB · tight
best quantization that fits · calculator engine
Spec
- ·AMD Threadripper Pro 7975WX (32C / 64T)
- ·256 GB DDR5 ECC RDIMM
- ·8 TB NVMe Gen 5 RAID 10
- ·2000 W redundant PSU
- ·Component manufacturer warranties
Runs for example
- ▸Llama 3.3 70B · Q8
- ▸Mistral Large 123B · Q5
- ▸Qwen 2.5 72B · Q8
- ▸Mixtral 8x22B · Q4
- ▸Llama 4 Scout · Q5
Software
- ·Everything in Pro +
- ·Multi-GPU tensor parallel configured (NVLink)
- ·Open WebUI multi-user · OIDC ready
- ·GDPR / AI Act compliance docs included
reference builds to assemble yourself · build cost varies with component market rates · Localia does not sell, does not ship
◆ custom
Beyond the reference builds.
Everything beyond the standard builds: full RAG integration on your documents, Grafana dashboards, custom SSO (fine-grained RBAC, HDS-compliant audit), white-label UI, datacenter GPUs (H100 / H200 / B200 / MI300X), multi-rack, strict SLA, security audit, team training. Localia does not sell these services; we can point you to the right partners and help you scope the need.
H100 / H200MI300XMulti-rackStrict SLAEnterprise RAGSecurity auditTeam trainingMaintenance
Ask us for advice →process
How to use these builds.
- 01Step 1
Choose a build
Find the reference build closest to your use case.
- 02Step 2
Estimate the cost
Open the calculator to estimate the build cost for your models.
- 03Step 3
Source the components
Buy the components yourself from the distributors of your choice.
- 04Step 4
Install the stack
Follow our guides to install Ollama, Open WebUI and the RAG stack.
faq
Frequently asked.
Does Localia sell these rigs?+
No. Localia is an independent resource: these builds are educational references to assemble your own rig. No stock, no sales, no shipping.
Which software stack should I use?+
Ubuntu 24.04 LTS · NVIDIA drivers · CUDA · Ollama · Open WebUI · llama.cpp · vLLM · RAG stack (LlamaIndex + Qdrant). All open-source, to install by following our guides.
Which models run on it?+
Depending on the build: Llama 3.3 70B, Mistral Large, Qwen 2.5 72B, DeepSeek R1, Gemma 4. The calculator tells you what fits in VRAM.
What about warranty?+
Each component keeps its own manufacturer warranty (GPU, motherboard, PSU, etc.). Localia is not a seller and therefore provides no warranty on the assembled whole.
What if I want to upgrade later?+
Every reference build is designed to accept +1 or +2 additional GPUs (free PCIe slots + sized PSU). Plan for it when buying the components.
Ready to run local?
A question about choosing a rig? Describe your case in two sentences: we'll point you to the right reference build — without selling you anything.
Ask us for advice →Reply within 24 business hours · contact@getlocalia.com