Boot a USB stick, and your AMD Ryzen AI MAX+ 395 (Strix Halo) machine serves local LLMs on its GPU and NPU, on Linux, through AMD's Lemonade Server. That is the whole setup.
- Download
1bit-os-<version>.img.zstfrom the releases page. - Write it to a USB stick (8 GB or more).
- Boot the machine from the stick. Secure Boot can stay on; the image is signed.
- Lemonade comes up with its catalog and web UI. Try it from the live session; to keep it, run
1bit-install <disk> --yesthere and the 1bit appliance is on the disk (os/v2/README.md).
Already running Lemonade on Ubuntu? The lemonade-onebit package from the same releases page adds the engine as a Lemonade backend; nothing else changes (docs/lemonade.md).
One binary to rule them all. The 1bit engine is model agnostic: any Hugging Face architecture, any GGUF quantization, Ryzen AI NPU packs and ONNX graphs, served by one Lemonade backend on the AMD Ryzen AI MAX+ 395 (Strix Halo). It runs the Radeon 8060S iGPU through HRX with our own HIP kernels, and the XDNA 2 NPU through our own NPU runtime, behind the OpenAI-compatible API Lemonade already gives you. Every model is checked on the hardware before it ships; the models and their measured numbers are one page: docs/models.md.
| when | what |
|---|---|
| 2026-10-04 | The engine becomes HRX + our HIP kernels + the NPU engine, embedded in Lemonade; Vulkan, ROCm and CUDA builds removed (RFC #213, #306) |
| 2026-10-20 | One HTTP hop from Lemonade to the engine's server; the stock Ubuntu lemond runs the engine with no patch (#475, #483) |
| 2026-10-21 | Every HRX pin change passes a hardware gate on Strix Halo before merge; Qwen3.8-27B decodes at 36 tok/s with MTP on our kernels (#478, docs/hw-gate.md) |
| 2026-10-23 | First release candidate built, booted from USB on the hardware with Secure Boot on (docs/releases.md) |
| 2026-10-25 | Microsoft BitNet b1.58 (TQ1_0 / TQ2_0) runs fully on HRX (docs/models.md) |
| 2026-10-30 | The HRX pin carries the DeepSeek-V4.1 port; ONNX Runtime runs in-process with our own execution provider on the NPU; the NPU and the GPU are measured running at the same time (#672, #647, docs/hybrid.md) |
| 2026-11-01 | v2026.44 is a draft: candidate 19 is gated in software (1959/1959 kernel ops, NPU 82.03 tok/s, VMTEST pass through install and installed-disk boot, VMTEST-RELEASE 10 pass / 9 skip / 0 fail) but has not been booted or installed on hardware; candidate 20 takes the full gate, hardware boot and install included, before it is published (docs/releases.md) |
The release goes out when every item on the release checklist is done (docs/releases.md); the weekly Lemonade pin moves with Lemonade's releases.
- 1bit.gg: the docs site, built from this repository's
docs/; the blog is the build log. - HRX on the Radeon 8060S and the NPU engine: the two device routes, on the site.
- docs/architecture.md: how the engine is built, what runs where, the hardware gate, the kernel, the direction.
- docs/serve.md:
1bit serveand its options. - Measured results: every number with how it was taken.
- Site repository: builds 1bit.gg from this repository.
The source code in this repository is Apache-2.0. See LICENSE.
The release binaries (the Linux package and 1bit OS) also contain two proprietary components, compiled in: the 1bit NPU runtime and the 1bit HIP kernels for HRX. They are not in this repository, their source is not published, and they are not under Apache-2.0: they ship under the 1bit Binary License (free to use for any purpose and to pass on unmodified; third-party parts in NOTICE-BINARY). A build from this repository without them has no NPU engine and runs HRX on the public kernels only (NOTICE, docs/releases.md).
Credit is welcome and, for the code, required: the Apache-2.0 license keeps the copyright notices and the NOTICE file with every copy. Three lines are enough:
1bit engine - https://github.com/1bit-MONSTER/engine - Copyright 2026 bong-water-water-bong - Apache-2.0
CITATION.cff carries the same in citation form (GitHub shows it under "Cite this repository"). If you build on a specific piece, name the file or the pull request it came from; the history is public.
Osmantic / ODS comes first. ODS introduced me to vibecoding, and that is where all of this started. Without it, this engine would not exist.
The Lemonade team and AMD's developers. Lemonade is where this engine runs. AMD's developers left breadcrumbs all over the place: the XDNA driver and XRT, IRON and Peano, HRX, their tested llama.cpp integration, their issues, their examples. This engine is what following those breadcrumbs built.
The repositories this engine is built on, in order of importance:
| # | Repository | What it gives this engine | License |
|---|---|---|---|
| 1 | amd/xdna-driver | The XDNA 2 NPU kernel driver (amdxdna) the NPU engine runs on |
GPL-2.0 (driver) |
| 2 | Xilinx/XRT | AMD's NPU runtime and its API, where the engine's NPU work started | Apache-2.0 (userspace) |
| 3 | Xilinx/mlir-aie | IRON and aiecc: how our own NPU kernels are written and compiled | Apache-2.0 WITH LLVM-exception |
| 4 | Xilinx/llvm-aie | Peano, the C++ compiler for the NPU's AI Engine cores | Apache-2.0 WITH LLVM-exception |
| 5 | ggml-org/llama.cpp | GGUF inference on the Radeon iGPU through HRX (AMD's tested pair), and on the CPU | MIT |
| 6 | ROCm/hrx-system (our fork 1bit-MONSTER/hrx-system carries two commits) | HRX, AMD's HIP Runtime Extended, behind the HRX0 device | |
| 7 | torvalds/linux | The kernel, with amdxdna and amdgpu in-tree |
GPL-2.0 WITH Linux-syscall-note |
| 8 | huggingface/tokenizers | Every model's tokenizer.json, byte-exact, behind our C ABI |
Apache-2.0 |
| 9 | NandhaKishorM/laya | The router that decides where each request runs | Apache-2.0 |
| 10 | ROCm/FastFlowLM | The Q4NX NPU model format and its models on Hugging Face (FastFlowLM/*-NPU2), which the engine's NPU route runs on its own kernels |
MIT |
| 11 | antirez/ds4 (DwarfStar) | reference only, not in the shipped build since v2026.44: antirez's CUDA kernels through a HIP shim plus hipBLAS are not our HIP kernels. The fork 1bit-MONSTER/ds4 stays as the reference for the work (dwarfstar.md) | MIT |
Also built on nlohmann/json (MIT) and yhirose/cpp-httplib (MIT).
Every third-party copyright and license is listed in NOTICE. Each project keeps its own license; nothing here relicenses anyone's work.