Home

LexiSuite: a complete local AI platform, in Rust, from the kernel up

LexiSuite is everything needed to run AI models on hardware you own, built as one system: an operating system that boots into a control panel, an inference engine that runs the models on your GPU, a kernel designed around inference, and the Panel that operates all of it. It boots from one USB image and nothing in it depends on a cloud service.

It is not a front end for someone else's engine and not a wrapper around llama.cpp. The engine, its GPU kernels and the Panel are original Rust code, 1.85 million lines in one codebase, developed daily by its author as their own primary AI platform.

Component	What it does
LexiEngine	Runs GGUF models on Vulkan (AMD, Intel, NVIDIA), CUDA and CPU, behind OpenAI-compatible and Anthropic Messages-compatible APIs
LexiPanel	One browser GUI for models, instances, chat, benchmarks, tuning, training, users, storage, fleets and virtual machines
LexiOS	Bootable appliance OS: startup, network, firewall, disk growth, recovery and signed updates
LexiKernel	A Rust kernel with a Linux application interface, built around inference; the second boot choice on the same stick

Why nothing else is like it

Local AI is usually assembled from separate projects: a Linux distribution, an inference server, a chat front end, a hypervisor, quantization and training scripts, monitoring. LexiSuite is all of those layers as one product, written to one set of rules and tested together.

• An engine measured against llama.cpp on every change. On a Radeon RX 7900 XTX on November 10, 2026, against llama.cpp b12230: a 27B model at a 240,128-token context generated faster at every depth measured, and LexiEngine completed 138,000-token and 231,000-token prompts where llama.cpp lost its Vulkan device three times out of three. A 9B model at a 524,288-token context generated faster at every depth, 48% faster at 509,000 tokens deep, where it also processed the prompt almost twice as fast. Full tables, and the cases where llama.cpp is still ahead, are on Benchmarks.
• One GUI for the whole machine. 1,158 features and 58 guided tasks in one Panel, from "serve your first model" to "boot an installer with GPU passthrough".
• Local by design. Prompts, models and data stay on the machine.
• One image, two kernels. The development USB image carries LexiOS on Linux 7.4 and LexiKernel side by side, with Secure Boot.
• Security that is tested. About 4,600 of the 15,607 tests exist to prove that something is refused.
• Honest status. It is in active development, not a general release; every page says what has been proven and what has not.

Start here

1. Getting Started: download the development image, verify it and choose a boot path.
2. Install from USB: flash with a graphical writer or dd, without overwriting the wrong drive.
3. Secure Boot: enroll the development certificate by appending to the firmware allow-list.
4. First Boot: locate the Panel and create the administrator account immediately.
5. Hardware and Kernels: select the best-supported environment.

New on 11/11/2026, each with its own status note: Write a stick from Linux, macOS or Windows; install to an internal disk; native installers for Linux, macOS and Windows; updating in place; ternary models (experimental).

Already running? Models → Instances → LexiChat. Then explore Benchmarks, security, troubleshooting and the FAQ.

One Panel interface

After first-boot enrollment, the Panel opens at /. Its tabs run across the top (Status, Models, Run, Optimize, Fleet, Files, System, Help); each tab opens a drop-down menu of its pages, and the open tab shows them as a row of sub-tabs. Every page opens inside that frame. Help → Find a feature searches every control, setting and guided task by name without changing anything. Start with Panel Overview, then Navigation and Help.

What can I do in the Panel?

Goal	Guide
Add and verify a model, including offline import and verified-only admission	Models
Start, stop, drain, profile or optimize an instance; choose its engine and devices	Instances, Optimizer, Builds
Inspect GPU/CPU, tune power, estimate capacity and compare engines	Power and hardware, Benchmarks and evals, Fit Lab
Evaluate, train, export, or use typed decisions and token-gating	Training, Decisions, TokenGate
Manage network, access, backups, updates, VMs, agents and diagnostic evidence	Network and access, Storage, Virtual machines, Agents, Diagnostics
Use interoperable API clients	Panel API, OpenAI-compatible API, Anthropic Messages-compatible API, Typed decisions API

Not all functions are available under LexiKernel; the Panel reports unsupported functions rather than silently claiming they worked. See Kernels.

Development status — November 10, 2026

This is a development/testing image, not a generally available or production-qualified release. Secure Boot was qualified in a virtual machine, not yet on a physical Secure Boot PC. LexiKernel has booted a real Intel laptop and served the Panel, and its own RDNA3 driver now starts a physical RX 7900 XTX up to a command buffer with an interrupt-delivered fence, but runs no shader yet; GPU operation there relies on a Linux driver-domain guest. CPU and storage-usage reporting on LexiKernel remain incomplete. One of five recent LexiKernel browser boots froze after its console appeared.

The USB image contains no models. It needs at least an 8 GB USB device; writing it erases that device. For installation and release changes, consult Release Notes; for the announced terms and an important difference between the published summary and repository licence, read Licence.

Reference controls

Control or choice	What it does
Boot menu: 1 LexiKernel	Starts the experimental kernel; requires supported IOMMU, USB and networking.
Boot menu: 2 LexiOS on Linux 7.4	Starts the current development image's Linux 7.4-rc2 entry; use it for broader hardware support.
First-boot setup → Complete first-boot setup	Creates the administrator account and records setup; do this before exposing the appliance to an untrusted network.
Models / Instances	Imports models and configures a running inference service; the image itself ships no model weights.

Home · Getting Started · Security · FAQ