Run frontier LLMs on your own hardware.

Yes. Open-weight models (Llama, Qwen, DeepSeek, Gemma, GPT-OSS) run entirely inside your own network on a single 128 GB unified-memory box such as an AMD Strix Halo or NVIDIA DGX Spark — no data leaves your perimeter, no per-token API bill.

On-prem removes the hardest PDPA risks — cross-border transfer and third-party processing — because data never leaves your environment. It is not automatically compliant on its own; you still need access control, audit trails and governance, which we build in. No serious vendor should claim 'fully PDPA-compliant' out of the box.

Measured live on our own AMD Strix Halo box, a 35B Qwen3.6 MoE model with native multi-token prediction decodes at around 65 tokens/second on chat and prose, and up to ~100 tokens/second on predictable code output, where speculative decoding accepts more draft tokens. It processes prompts at roughly 840 tokens/second. A NVIDIA DGX Spark is about 25% faster again. The newer dense Qwen3.8-27B, a vision-native model, decodes at a measured 20-22 tokens/second on the same box with multi-token prediction, and its hybrid linear-attention design keeps long context cheap: we serve two 131k-token slots concurrently on one machine. That is comfortably interactive for chat, extraction and agent workloads.

For a single team, a 128 GB unified-memory box: AMD Strix Halo (Ryzen AI Max+ 395) for best value, or NVIDIA DGX Spark for the CUDA ecosystem. Both hold a 100B-class MoE model. For many concurrent users you move to data-centre GPUs (A100/H100), which trade cost for throughput.

AMD Strix Halo gives the most memory per dollar and runs the largest models on a desk via Vulkan/RADV; NVIDIA DGX Spark is ~25% faster, steadier per-token, and has the mature CUDA/vLLM/TensorRT tooling. Pick AMD for cost-per-token, NVIDIA for tooling and mixed workloads.

When data can't leave your perimeter (regulated data), when you sustain enough volume that per-token API costs outrun hardware (typically millions of tokens/day), or when you need predictable low latency. Below that, cloud APIs are usually cheaper and faster to start — we'll tell you honestly which side you're on.

Yes. A 128 GB unified-memory box holds a 100B-class model in 4-bit — we run a 122B MoE unit locally. Mixture-of-Experts models are the sweet spot: large quality, only 3–10B parameters active per token, so they stay fast on memory-bandwidth-bound hardware.

Yes, and it is how we run our own. A 128 GB box holds a fast MoE drafter (Qwen3.6-35B, ~3B active parameters, ~65 tokens/second) alongside a smarter dense vision model (Qwen3.8-27B, ~20 tokens/second, reads images and documents natively), each serving a different port. Quick chat and extraction route to the fast one; harder reasoning, coding and anything with images routes to the smart one. Same hardware, no cloud, both measured on our own machine.

A working proof of concept in 1–2 weeks and daily production use in 6–8 weeks is typical — hardware sizing, model selection and quantization, the inference gateway, governance and audit, then integration with your systems. We hand over a documented, self-serve stack, not a black box.

Llama, Qwen (incl. Qwen3.6 MoE and the vision-native Qwen3.8), DeepSeek, Google Gemma, GPT-OSS and Mistral families, and writing-tuned models such as Muse Glimmer, plus OCR/vision models (Qwen-VL, Surya) for document workloads. We match the model to your task, hardware and quality bar rather than defaulting to the biggest one.

Either. We advise on and size the hardware (Strix Halo, DGX Spark, Apple Silicon, data-centre GPUs), or deploy onto kit you already own or run in your own VPC. The whole stack — models, gateway, governance — runs inside your control; day-to-day ops stay with your team unless you want a managed retainer.

It's our packaged private-LLM deployment: a complete stack — model server, authenticated gateway and secure access — on your own hardware, shipped with Deneb, our open-source setup agent that installs, diagnoses and repairs the stack on the box, with a human engineer behind it. The client is auditable line by line, read-only by default, and never runs anything destructive or privileged. Your data and model never leave your walls.

We red-team your AI features with Vedis, our AI-security appliance — prompt injection, multi-turn jailbreaks, indirect injection via RAG/tools, data-leakage and tool-abuse — mapped to the OWASP LLM Top 10. It runs entirely on your own hardware so nothing it tests leaves your network, and it's strictly defensive: authorized testing of your own apps, with a blue-team loop for scheduled re-scans and per-deploy regression checks.