LexiPanel is a Rust-based inference optimization and control plane for open models. It automatically optimizes quantization, selects the best runtime, schedules requests with KV-awareness, manages distributed deployment, and proves every change with paired statistical tests. Measurements are from real workloads on your hardware. Free for private, research and educational use · $100 per instance per year for commercial use.
The demo is the real 1.0.0 panel, running on a scrubbed snapshot of our development box: a Radeon 7900 XTX and an RTX 2060.
of identical settings are ever kept. A real 12% gain is kept 95% of the time on a thorough budget.
simulated at the box's measured noise, 4,000 trialsfaster decode for a 27B model on one 7900 XTX, from a smaller file at the same perplexity.
51.7 vs 46.5 tokens/s at a 91k contextrequests of a burst served through the KV-aware gateway. Sent straight to the server: 0 of 48.
RTX 2060, 8k shared KV poolfalse-win rate the platform measured for itself on this box, against 5% allowed.
Bench → Calibration, 6 blocksThe problem
Serving an open model means choosing a file format, an engine and backend, a speculative decoding method, batch sizes, cache types and a context length. Each choice is worth double-digit percentages, each depends on the others, and all of them move with every engine release. Most teams copy settings from a forum and never learn what they left on the table, or what quietly broke. Four things we measured on our own box:
N-gram speculative decoding is free speed.
On an RTX 2060 with a 1.7B model, ngram-simple wrote an edit at 473 tokens/s against 107 without it. On a 7900 XTX serving a 27B model with its own MTP head, n-gram drafting didn't help, and one mode ran a task past its 17-minute timeout.
ROCm is faster than Vulkan on AMD.
At a 91k-token context on a 7900 XTX, AMD's ROCm 10.1 build read prompts 50% faster and wrote 22% slower. Weighted by the box's real traffic (48% reading, 52% writing), neither wins, and only Vulkan held the 245,760-token context with MTP.
FP4 is the new 4-bit.
On a model not trained for it, MXFP4 cost +17.1% perplexity against +1.8% for Q4_K_M and +2.1% for IQ4_XS, and decoded no faster than IQ4_XS on either backend.
The bigger file is the better model.
Three quantizations of the same 27B at a 91k context: the smallest (15.3 GB) decoded 51.7 tokens/s against 46.5 for the largest (18.5 GB), with a perplexity gap a tenth of its error bar.
The right answer depends on your hardware, your model and your traffic. So LexiPanel doesn't assume. It measures, decides on evidence, and keeps measuring.
How it works
One run's numbers never change a setting. Every change goes through the same loop, and every step is recorded: which file (by SHA-256), which build, which settings, which numbers.
Every GPU with its VRAM, power and temperatures; the model from its file header; the real workload: context depths, prompt and answer sizes, speculative acceptance.
Settings to change, files to try (per-tensor requants, other engines' formats) and devices to place them on.
Screen each candidate once, then confirm the best in paired, interleaved A-B/B-A rounds. It's kept only if the whole 95% interval clears the margin and the answers aren't worse.
Drain, hand over the KV cache, restart, verify with a canary. A configuration that doesn't come up is rolled back.
Judge the requests that follow against the speed expected at their own depth, and roll back when they got slower.
Starting from a deliberately slow micro-batch, the optimizer kept one candidate, audited its settings' neighbours with the others in place, then validated the combination.
RTX 2060 · Qwen3-1.7B Q8_0 · 16k context · paired intervals of 1–2%, 4–8 pairs each · saved settings restored at the end
The product
21 tabs, from the first model download to the GPU's fan curve. Every screen below is from the live demo, which you can open and click through.
A real run from today, on the 7900 XTX, searching the speculative strategy for a 27B model.
What your requests actually look like, learned from the traffic you already serve.
Per-tensor requantization fitted to your cards, starting from what each format actually does on them.
Clocks, voltage, power cap and fan curve through a range-checked helper, with every change measured.
Throughput straight from the engine, request by request.
Measurements that say how sure they are, starting with the platform's own error rates.
Engines, inference fabric, and Quant Forge
LexiPanel 1.1.0 adds a Rust-based inference fabric with KV-aware scheduling, SLO-aware routing, and P/D disaggregation. Quant Forge generates optimized GGUF artifacts with exact tensor allocation, per-tensor precision targeting, and BPW/MiB budgets. One control plane unifies quantization, runtime selection, scheduling and distributed deployment.
KV-aware routing, SLO-aware scheduling, auto-scaling across devices and fleets.
Trust
A release fails its own audit when a documented capability has no test behind it, a route doesn't exist, or versions disagree. In the 16 October 2026 software gate, 7,712 tests pass and 0 fail: 1,716 Panel, 3,877 Engine, 2,119 LexiOS, with Vulkan validation 360/0.
1,186,842 lines of Rust across 2,580 files, 95.89% of all source. Clippy runs clean on Panel, Engine CPU and GPU, and LexiOS. Engines install into their own environments.
Roles and API keys, a hash-chained audit log, CSRF and DNS-rebinding guards, login backoff. No unquoted value ever reaches a shell, and a test on the source enforces it.
Journaled, drained, verified with a canary, and rolled back when a new configuration doesn't come up.
51 MCP tools let any MCP client read and operate the box, through the same guard rails and audit log. One switch makes them read-only.
Signed remote actions across boxes, with drain and restart holds. A box joins only when you let it.
Labels follow the source capability matrix. An implemented path means an identifiable API, UI or CLI entrypoint with real validation — not production support, hardware coverage or a signed release. Qualification gates still open: full-model parity, performance, CUDA serving, 24-hour soak, clean install/update/rollback, production signing and physical LexiOS ratification.
Why now
A 27B open model with its own speculative head decodes 59 tokens/s at an 8k context and 47 at 32k on one 24 GB consumer GPU. We measured it on a 7900 XTX.
New engines, formats (FP8, FP4, IQ quants), speculative methods (MTP, EAGLE-3, DFlash, n-gram) and backends. Each release moves the best configuration.
Engines ship knobs, and benchmarks measure someone else's machine. LexiPanel decides on your hardware with your traffic, and keeps the evidence.
The suite
The web control panel: model library and downloads, engine instances per GPU or CPU, parameters and templates, benchmarking and auto-tuning, Fit re-quantization for the cards a machine actually has, GPU tuning and power caps, statistics, logs, diagnostics and crash evidence, a file manager, built-in chat, OpenAI-compatible and Anthropic-Messages-compatible API endpoints, accounts and API tokens, fleet management, a System tab, a hardware security check, and a complete console CLI.
A Linux-based appliance OS on the Linux 7.3 kernel that boots from a USB stick straight into LexiPanel. First-boot setup in a browser or on the machine's own console, works with no network at all through link-local addressing, installs to an internal disk from the Panel, A/B updates and recovery, signed UKI boot under Secure Boot with a verified read-only system (opt-in in this build), a TPM-signed boot report, and a hardware and firmware vulnerability check at install. First real-hardware tests are in progress.
Bundled as the default engine, as an upstream build in a verified bundle.
Our own Rust inference engine. It runs Qwen3 on CPU, AMD and NVIDIA GPUs today, and is slower than llama.cpp. The Gemma 4 and DeepSeek-V2 GPU paths are new and not yet validated on GPUs.
An experimental Rust kernel aimed at AI-first appliances. It boots in virtual machines and runs the Panel; it is not yet ready for real hardware.
Status and evidence
Build 0c335da, a LexiOS image under QEMU/KVM, in virtual machines with real GPUs passed through using VFIO. Model Qwen3-1.7B (Q4_K_M), one short chat prompt through the Panel's gateway. The times include loading the model.
| Device | Engine | Result | Chat step |
|---|---|---|---|
| NVIDIA GeForce RTX 2060 (Vulkan, open NVK driver) | LexiEngine | PASS | 37.8 s |
| NVIDIA GeForce RTX 2060 (Vulkan, open NVK driver) | llama.cpp | PASS | 11.7 s |
| CPU only | LexiEngine | PASS | 103.3 s |
| CPU only | llama.cpp | PASS | 28.8 s |
| AMD Radeon RX 7900 XTX (Vulkan) | LexiEngine | PASS | 31.2 s |
| AMD Radeon RX 7900 XTX (Vulkan) | llama.cpp | PASS | 2.4 s |
AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 2060, and CPU-only.
Vulkan GPUs from AMD, NVIDIA through the open NVK driver, and Intel; CUDA on NVIDIA when compiled for it.
An x86-64 CPU with AVX2, 8 GB of RAM (16 GB or more for larger models), UEFI firmware, and Ethernet for network setup — Wi-Fi setup is not available yet. CPUs without AVX2 are detected and reported; inference on them is unsupported or very slow.
About 1.7 million lines of Rust across the Panel, Engine, OS and kernel. 99.6% of production code is safe Rust.
Where it goes next
Support the project
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.
Three full packs of Black Pine Genetics seed, selected at random from current stock and shipped to you.
Five full packs, selected at random from current stock and shipped to you. The best value per pack.
Ten full packs, selected at random from current stock and shipped to you. The deepest support tier.
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.
Licensing and deployment
Free for private, research and educational use. Commercial use is licensed per instance, at $100 per instance per year. White-label — your brand on the Panel and the OS — is available under the commercial licence. Deploy on-prem, air-gapped, or in your own cloud. The source code will be published as source-available under a non-commercial licence; the exact licence text is being finalized. Contact us for commercial licensing, white-label or volume terms.