Fully Local AI Deployment
Open-weight models from Canada, Europe and beyond, including Cohere’s Command A+, Mistral Large 4 and Ripple-65B, selected for each workload and running on computers you own. Deployment runs in a private cloud, on your own hardware, or fully air-gapped, supporting Canadian sovereign AI and reducing dependence on foreign AI clouds.Open-weight models we have tested ourselves, running on computers you own: private cloud, on-premises or fully air-gapped.
Using a closed cloud AI service sends prompts, code, and designs outside your environment. Retention and training terms vary by provider, plan, and contract. Local or air-gapped deployment keeps proprietary work from being transmitted to an external model provider. With models running on Canadian-controlled infrastructure, organizations can also keep more of their data, operations, and AI capability under Canadian control.Cloud AI sends your prompts, code and designs outside your environment. Local deployment keeps them in.
You Do Not Need a Data Center to Run Local AI
AI does not always require a new building, a new power connection, or a massive cooling system. For many businesses, local models can run on a workstation or compact server in the space they already have. The less a deployment relies on data center capacity, the smaller its environmental footprint. Keeping equipment in service longer reduces it further.Many businesses can run local AI on a workstation or compact server in space they already have.
Less Land, Concrete, and Grid Pressure
A large AI data center brings a large construction footprint: land clearing, concrete and steel, cooling equipment, new transmission capacity, and continuous electricity demand. Those impacts arrive before the first prompt is processed. If your workload can run locally, you do not need to add that infrastructure.A new AI data center means land, concrete, cooling and grid capacity. A workload that runs locally needs none of it.
Local AI Can Fit in the Room
A compact AI workstation such as NVIDIA DGX Spark can plug into an ordinary wall socket and run capable open models close to the people using them. It needs a desk, not a new building. That means less construction, no land cleared for your deployment, and a smaller physical footprint than a dedicated data-center build.A compact workstation like NVIDIA DGX Spark plugs into a wall socket and runs capable open models. It needs a desk, not a building.
Work That Never Needs a Data Center
Every workload that runs on your own equipment is one that does not need capacity built for it somewhere else: no land, no cooling system, no new connection to the grid. It runs on the power and space you already have. Hardware kept in service longer, and recycled responsibly at the end, keeps the footprint smaller still.Work on your own equipment needs no new land, cooling or grid connection, just the power and space you have.
Open Weights from Canada, Europe and Beyond
Ripple-65B runs on our own AMD workstation; the measured speeds are on our test results page. Cohere’s Command models are built in Toronto and Mistral Large 4 in Paris. Running open weights on Canadian-controlled infrastructure keeps your data and operating continuity under your control.Open weights from Canada, Europe and beyond, on infrastructure you control.

Command A+ & North Mini Code
Cohere has built the Command family of enterprise models in Toronto since 2019. Command A+ is its largest open model under Apache 2.0: a 218B mixture of experts with 25B active per token, a 128K context and 48 languages. North Mini Code, its agentic coding model, is far smaller at 30B with 3B active and a 256K context.Built in Toronto: Command A+ and North Mini Code, both under Apache 2.0.

Mistral Large 4
Mistral Large 4, released in October 2026 by Paris-based Mistral AI, is a mixture of experts with about 1 trillion parameters, 49 billion active per token. It reads text and images, holds a 1M-token context and is fluent in more than 160 languages, including every official EU language. Its MIT-based open-weight license allows commercial use; only reselling it as a hosted service needs a separate license.Built in Paris: a 1-trillion-parameter mixture of experts fluent in 160+ languages.

Ripple-65B
Ripple-65B Preview is a dense 65-billion-parameter model from Ricursive Intelligence in Palo Alto, built for agentic coding and reasoning, under the Apache 2.0 license. On our Radeon build its 35 GB 4-bit version runs entirely in GPU memory, split across both cards by vLLM, at 28 tokens per second with a 64K context.Apache 2.0 dense model for agentic coding, 28 tokens/s in GPU memory on our Radeon build.
Run on Your Own Hardware
All-in-one AI workstations, inference servers we build to order in Canada, or hardware you already own, sized for your workload, number of users, and performance requirements.
Servers & Workstations
We can use hardware you already own, including compatible memory and capable business machines, or specify a new workstation or server. Each vendor has its own acceleration path: MLX on Apple Silicon, CUDA on NVIDIA, ROCm on AMD, oneAPI on Intel, with Vulkan across all of them when portability matters more. Deployments range from one workstation to multi-GPU servers for a whole team.We reuse hardware you already own or specify new, from one workstation to multi-GPU servers, on Apple, NVIDIA, AMD or Intel.
Apple Mac, AMD Ryzen AI Halo, NVIDIA DGX Spark
Complete workstations built to each vendor’s own first-party reference design: Apple Silicon Macs, AMD Ryzen AI Halo systems, and NVIDIA DGX Spark GB10 or GB300 workstations. The platform is matched to model size, user count, and the performance your workload needs.Reference-design workstations from Apple, AMD and NVIDIA, matched to your model size and number of users.
Custom-Built Inference Servers
We build inference servers to order in Canada, choosing the accelerator to match the job: NVIDIA RTX PRO 6000, AMD Radeon AI PRO R9700, or Intel Arc Pro B70. Retired data-center inference cards are another option. They usually have years of useful work left, cost well below new, and put no fresh demand on a memory market where prices are already high.Inference servers built to order in Canada with NVIDIA, AMD or Intel accelerators, or retired data-center cards that cost far less than new.
Inference Runtimes & Model Management
Inference runtimes execute and serve the models. Desktop and web apps make it easier to run, fine-tune, and build generation workflows with them.
Linux, Windows & macOS
Production inference can run on Linux, Windows, or macOS, from an Apple Silicon workstation to bare-metal and virtualized servers. Fully offline, air-gapped configurations are available when data must remain inside your environment.
vLLM, SGLang and llama.cpp
vLLM and SGLang serve high-throughput team workloads, while llama.cpp runs quantized models across common hardware. We select the runtime for the model, operating system, hardware, and number of users.vLLM and SGLang for teams, llama.cpp for common hardware: chosen for your model and users.
Unsloth Studio & ComfyUI
Unsloth Studio runs language models and fine-tunes them on your own documents, then exports them to llama.cpp or vLLM. ComfyUI builds image, video and audio generation from connected nodes. Both are open source and run on your own hardware.Unsloth Studio to run and fine-tune language models, ComfyUI for image, video and audio generation, all on your own hardware.
Maya & Strata: Big Models from an SSD
Mixture-of-experts models use only a small share of their weights for each token. Two new open-source engines, Strata and Maya, take advantage of this: they cache the busiest experts on the GPUs, keep a working set in system memory, and read the rest from a fast NVMe SSD only when a token needs it. Models larger than a computer’s combined graphics and system memory can then run on consumer hardware.Strata and Maya stream model weights from a fast SSD, so models larger than the computer’s memory run on consumer hardware.
GPU, Memory and SSD Tiers
Each token activates only a few experts. The most-used experts stay cached in GPU memory, a bounded working set sits in system RAM, and the rest are read from the NVMe SSD on demand. The context cache is compressed and streamed, and both engines split the model’s layers across both cards. Memory guards keep the model out of swap and leave room for the rest of the computer.Busy experts on the GPUs, a working set in RAM, the rest read from the SSD as needed.
Qwen3.8-Flash-Next at 256K Context
Strata runs the 76 GB IQ3_XXS build of Qwen3.8-Flash-Next with its full 256K context on our two 24 GB Radeon cards and a 128K context on our GeForce build, streaming expert and per-layer data from the SSD. On the GeForce build it reaches 84 tokens per second and 1,058 tokens per second of prefill. Strata also runs DeepSeek V4.1 Flash, at 12 tokens per second on the GeForce build.A 76 GB model with a 256K context, and 84 tokens/s on our GeForce build. Strata on GitHub →
DeepSeek V4.1 Flash on Consumer GPUs
Maya, a fork of Strata, runs DeepSeek-V4.1-Flash-1017 and GLM-5.3-Flash. On our GeForce build, Unsloth’s 233 GB DeepSeek IQ1_S build is about twice the machine’s 118 GB of combined graphics and system memory, yet it generates 8 tokens per second with a 32K context. The 95 GB GLM build reaches 15 with a 64K context.A 233 GB DeepSeek model on a 118 GB machine, at 8 tokens/s. Maya on GitHub →
Agentic Workflows
Turn local models into practical tools and controlled automation: private chat and knowledge, coding and business agents, and secured autonomous workflows.
Private Chat & Knowledge
Open WebUI gives your team a familiar multi-user chat with per-user access control. It can connect local models to approved documents, code repositories, vector databases, business tools, and private memory without sending that context to an external model provider.
Coding & Business Agents
OpenAI-compatible APIs let OpenClaw, Hermes agents, Claude Code, Codex, and internal business agents use models running on your hardware. Existing workflows can move from hosted AI to local models without rebuilding the whole toolchain.OpenAI-compatible APIs let Claude Code, Codex and your own agents use local models without rebuilding your tools.
Controlled Autonomous Work
We harden agent pipelines with prompt-injection defenses, least-privilege tool permissions, sandboxing and allowlists, audit logs, and human approval gates for consequential actions. Agents can automate useful work without receiving unrestricted access to everything around them.Prompt-injection defenses, least-privilege permissions, sandboxing, audit logs and human approval for consequential actions.
Real-World Test Results
We measure every model we recommend on our own workstations, a dual Radeon RX 7900 XTX build and an RTX 3080 Ti + RTX 3080 build, and publish the speeds, context lengths and settings.Measured on our own AMD and NVIDIA workstations.
Ready to get started?
Tell us about your environment, we’ll design a deployment that keeps your data yours.


