Policy, infrastructure and model must each be answered. A local data centre without control of the weights is residency without sovereignty.
Hardware fit
What runs inside your perimeter
Pick a representative machine and a capability class, or choose Other to enter your own VRAM, RAM and CPU. The table ranks open-weight examples by estimated memory fit for architecture conversations, not a live probe of your rack.
GPUNVIDIA A100 80GBVRAM80 GBRAM256 GBCPU64 cores
LLM instruct weights. Showing 9 models for this profile.
Model
Size
Fit
Mem
Quant
Mode
Est. tok/s
Qwen2.5-3B-Instruct
Qwen · Apache-2.0
3B
Perfect
2.8 GB · 3.6%
Q5_K_M
GPU
120
Mistral-7B-Instruct-v0.3
Mistral · Apache-2.0
7B
Perfect
5.7 GB · 7.3%
Q5_K_M
GPU
92.6
Qwen2.5-7B-Instruct
Qwen · Apache-2.0
7B
Perfect
5.9 GB · 7.5%
Q5_K_M
GPU
89.5
Mistral-Nemo-Instruct-2407
Mistral · Apache-2.0
12B
Perfect
9.4 GB · 12%
Q5_K_M
GPU
56.2
OLMo-2-1124-13B-Instruct
AllenAI · Apache-2.0
13B
Perfect
10.5 GB · 13.4%
Q5_K_M
GPU
50.3
Qwen2.5-14B-Instruct
Qwen · Apache-2.0
14B
Perfect
11.2 GB · 14.3%
Q5_K_M
GPU
47.1
Qwen2.5-32B-Instruct
Qwen · Apache-2.0
32B
Perfect
20.5 GB · 26.1%
Q4_K_M
GPU
25.8
Qwen2.5-Coder-32B-Instruct
Qwen · Apache-2.0
32B
Perfect
20.5 GB · 26.1%
Q4_K_M
GPU
25.8
Qwen2.5-72B-Instruct
Qwen · Apache-2.0
72B
Good
42 GB · 53.5%
Q4_K_M
GPU
12.6
Estimates use a reserved memory band and representative quants. Confirm on an offline demo, cable pulled , on your own documents and drivers.
Customer is controller and processor. We license software, never hold production data.
Infrastructure
Data in the customer data centre or in-region tenancy, air-gapped, private cloud or bare metal.
Model
Customer controls the weights. Open weights on their hardware; licence files on disk.
What's in the box
Vision/OCR, LLM, embeddings and reranking, Apache 2.0 or MIT only. Named list on Proof →
Regulatory mapping
For health insurers, TPAs, financial institutions and government-adjacent entities, public cloud AI is often a compliance blocker, not a preference. In an offline demo we map who is controller, where records rest, who holds keys, which licence files sit on disk, and how audit is written without egress.
The stack enables those answers with controls in the path: customer perimeter and keys, Model Gateway, policy and DLP before weights, RBAC, human-in-the-loop on consequential actions, and immutable audit on the same install. Buyer regimes such as Health ICT rules, CBUAE outsourcing, and PDPL where in scope are why those controls exist. We describe vendor categories, not named competitors.