Sheet 01Version ae24156 · 11/16/2026 · all Rust

Local AI is an archipelago. This is the continent.

LexiSuite is everything needed to run AI models on hardware you own, built as one system: an inference engine measured faster than llama.cpp on large models and long contexts, the control panel that runs the whole machine, an OS that boots from a USB stick, and a kernel designed around inference. One codebase. Nothing in the cloud.

Free for private, research and educational use · $100 per instance per year commercially

Sheet 02Measured, not claimed

Watch it outrun llama.cpp and Strata.

Both lanes stream text at the speed each engine measured on the same Radeon RX 7900 XTX, with the same model file and the same settings, on 11/16/2026 against llama.cpp 1.0.0, and on Qwen3.8 Flash-Next against Strata 0.2.13. Pick a case. We lose the RTX 2060 case.

LexiEngine
generation after the first token
llama.cpp
Vulkan, flash attention
262ktokens of context answered by a 27B model on one 24 GB card: the model's whole window. llama.cpp 1.0.0 lost the device on a 134k prompt and returned nothing beyond.
+45%generation ahead of llama.cpp 1.0.0 at 253,000 tokens deep on a 9B model, and the prompt read more than twice as fast.
1.95Mlines of Rust in 340 crates: kernel, OS, engine and Panel in one codebase.
16,553tests. About 4,200 of them exist to prove that something is refused.

Qwen3.8 Flash-Next against Strata

A 94 GB mixture-of-experts model on one 24 GB Radeon RX 7900 XTX at a 262,144-token context with a 4-bit cache, 11/16/2026. Strata 0.2.13 ran the same token ids on the same card and files with its recommended settings. The better figure in each pair is in bold.

Prompt tokensFirst token, LexiEngineFirst token, StrataGeneration, LexiEngineGeneration, Strata
15,99839.3 s65.4 s4.68 tok/s2.43 tok/s
31,99774.2 s111.8 s11.36 tok/s5.75 tok/s
63,988151.2 s215.4 s3.03 tok/s2.96 tok/s
127,988322.2 s409.9 s5.26 tok/s3.06 tok/s
255,999711.7 s804.1 s3.76 tok/s4.30 tok/s

The generation figure at 255,999 tokens is from the engine’s runner. Method and every other table: Benchmarks.

Sheet 03Why nothing else is like it

Every layer, one maker.

Running models on your own hardware normally means assembling an inference server, a chat front end, a Linux distribution, a hypervisor, quantization and training scripts and monitoring, each from a different project. You are the ferry between the islands. LexiSuite is every one of those layers, built to the same rules, tested together and shipped as one image.

LayerNormally assembled fromIn LexiSuite
Inference enginellama.cpp, Ollama, vLLM, LM StudioLexiEngine. Original Rust engine with its own GPU kernels, not a wrapper. llama.cpp can be installed beside it from the Panel, so you can repeat every comparison yourself.
Control and chatOpen WebUI, a dashboard, SSHLexiPanel. One authenticated GUI for every operation on the machine, from the first model download to the GPU's power cap.
Operating systemUbuntu, Debian, ProxmoxLexiOS. Boots straight into the Panel. Firewall verified at boot, signed updates with rollback, a self-test on every start.
KernelLinux, general purposeLexiKernel. Rust, Linux ABI, inference-first scheduling. Linux 7.4 is the other boot entry on the same stick.
APIsWhatever the server offersOpenAI Chat, Completions and Responses, Anthropic Messages, embeddings, rerank, and typed decisions that return a probability for every option.
TuningHand-edited flags, forum postsBenchmarks against a reference engine, an optimizer that keeps only measured winners, and re-quantization to fit the card you have.
Training and VMsPython environments, QEMULoRA fine-tuning from the Panel, and a safe-Rust virtual machine monitor with GPU passthrough.

Sheet 04The platform

Four territories, one border.

Plate I

LexiEngine

Ahead of llama.cpp on large models

Our own Rust inference server: GGUF models on Vulkan (AMD, Intel, NVIDIA), CUDA and CPU. Speculation for greedy and sampled requests, contexts up to 1,048,576 tokens, no memory budget to set. Serves Qwen3, Qwen3.5 and 3.8, Gemma 4 and DeepSeek-V2.

Plate II

LexiPanel

Release candidate

Model library and downloads, instances per GPU, chat, benchmarking and auto-tuning, fitting, training, GPU tuning and power caps, statistics, crash evidence, files, a terminal, users and API keys, fleets and virtual machines. One page, tabs across the top.

Plate III

LexiOS

Release candidate

An appliance OS on Linux 7.4 that boots from a USB stick straight into the Panel. First-boot setup from a browser, Secure Boot with a development certificate, and a self-test that says what it could and could not prove.

Plate IV

LexiKernel

Boots real hardware

A Rust kernel with a Linux ABI: inference-first scheduling, VRAM admission, an IOMMU required by design. It has booted a real laptop from USB and served the Panel to another machine. Its own driver now starts a physical RX 7900 XTX as far as running a command buffer with an interrupt-delivered fence; GPU inference still goes through a Linux guest.

It is developed every day by its author as their own primary AI platform.

Sheet 05The Panel

The whole machine, in one page.

Every picture is the real Panel. Open the live demo and click through it yourself.

The Panel's Status tab: a 27-billion-parameter model serving on a Radeon 7900 XTX beside a second instance on an RTX 2060, with live throughput.
StatusWhat is running on which card, live.
The Optimize tab's results: measured candidates and the one that was kept.
OptimizeSettings are kept only when a measurement says so.
The GPU tuning tab: clocks, voltage and a power cap for the card.
GPUClocks, voltage and power caps per card.

Sheet 06Get there

A USB stick, a browser, a model.

Download

Take the image from the download page: one image of about 350 MB that carries the kernel, the OS, the engine and the Panel. No account, no licence key.

Write and boot

On Linux one command checks the image, writes it and reads it back: sudo bash install-to-usb.sh. On Windows or macOS use any image writer. Boot the PC from the stick.

Open the Panel

The machine's screen shows an address. Open it from any browser on your network, create the administrator, download a model and chat, or point an OpenAI or Anthropic client at it.

Sheet 07Surveyed and unsurveyed

What is proven, and what is not.

Proven

  • Ahead of llama.cpp 1.0.0 on a 27B Qwen model at a 262,144-token context at every depth measured (11/16/2026): generation 19% to 26% faster and prompts read 26% to 36% faster up to 33,000 tokens, and 134,000-token and 262,000-token prompts answered where llama.cpp 1.0.0 returned nothing at 134,000. A 32-turn agentic coding session on the same model: 197 s against 310 s. A 9B Qwen-family model at a 524,288-token context is ahead at every depth to 253,000 tokens: 18% to 45% faster in generation, and prompts read more than twice as fast at 253,000 tokens. Gemma 4 31B and E4B generate 4% to 16% faster at default settings, and the 31B answers a 201,000-token prompt with prompt processing 37% faster.
  • On Qwen3.8 Flash-Next, a 94 GB mixture-of-experts model served from one 24 GB card, it reaches the first token of a 16,000- to 256,000-token prompt in 11% to 40% less time than Strata on the same card: 39.3 s against 65.4 s at 15,998 tokens and 11 min 52 s against 13 min 24 s at 255,999 (11/16/2026, both through their own servers or runners on the same token ids).
  • The test set on the source of the current image (ae24156), 11/16/2026, on Rust 1.100.0: 4,199 engine, 570 server, 2,354 Panel, 1,616 LexiOS and 1,740 LexiKernel tests passed and none failed in those suites; and the 185 CUDA kernel comparisons on an RTX 2060 passed. The engine and Panel compiled for Windows and Apple silicon with no errors on 11/12/2026 (compile check only).
  • The USB image passed every boot of its virtual-machine qualification, including refusing a tampered kernel.
  • LexiKernel’s own driver started a physical RX 7900 XTX up to a command buffer whose completion arrives by interrupt.
  • LexiKernel’s own driver started a physical RTX 2060: the card’s firmware runs, a command channel is accepted, and its copy engine copied a page correctly (11/12/2026, bring-up build).
  • LexiKernel booted a real laptop from USB and served the Panel to another machine.
  • LexiKernel booted a Ryzen desktop (MSI X570, Ryzen 9 5900X) from USB on 11/12/2026: both NVMe drives and the 2.5G Realtek card bind, the Panel is served to another machine, and the USB keyboard works on the local console.
  • LexiEngine served a 27B model at a 200,000-token context to an agent application on another machine, tool calls included.

Not yet

  • RTX 2060: llama.cpp 1.0.0 generates 4% to 15% faster on a 9B model and processes prompts about 3 times as fast. Gemma 4 31B at its largest contexts: a later prompt on the same server can be refused for memory. Qwen3.8 Flash-Next: the draft head is not used in serving yet.
  • The Secure Boot image on a physical Secure Boot PC.
  • GPU inference on LexiKernel without its Linux driver guest: its own XTX driver has not run a shader yet.
  • The RTX 2060 under LexiKernel beyond a copy job: no graphics or compute context, no shader, no model on that card yet.
  • Windows and macOS: nothing has been linked or run there yet.

Support the project

Keep it free for everyone else.

LexiPanel is free for private, research and educational use, and it stays that way. Supporter packs fund the work — and every one of them ships you real seed from Black Pine Genetics.

Supporter

3 random seed packs

Three full packs of Black Pine Genetics seed, selected at random from current stock and shipped to you.

Supporter · Plus

5 random seed packs

Five full packs, selected at random from current stock and shipped to you. The best value per pack.

Supporter · Patron

10 random seed packs

Ten full packs, selected at random from current stock and shipped to you. The deepest support tier.

Commercial licence

Per instance, per year

Required for commercial deployment. Free for private, research and educational use — nothing to buy and nothing to register. There is no licence key and no activation: licensing runs on the honour system. Includes no physical goods.

Supporter packs include physical goods shipped from Black Pine Genetics and are available to adults 21 and over only, in jurisdictions where such seed may lawfully be sold and received. Seed selection is random and cannot be chosen.

Sheet 08Licence

Free to use. Fair to sell on.

Free for private, research and educational use. Commercial use is $100 per instance per year, which includes support and priority fixes for problems specific to your hardware. No licence keys and no usage checks: commercial users are trusted to buy what they use. White-label is available under the commercial licence. The source code is proprietary.

Private, research, educationFree
Commercial$100 / instance / year
Licence keys, activationNone
RunsOn-prem, air-gapped, your cloud