1bit engine: local LLM inference on AMD Strix Halo (Ryzen AI MAX+ 395)

One binary to rule them all: the 1bit engine is a model-agnostic local LLM inference engine for the AMD Ryzen AI MAX+ 395 (Strix Halo) on Linux. It serves any GGUF on the Radeon 8060S iGPU (gfx1151) through HRX with our own HIP kernels, and models on the XDNA 2 NPU through our own NPU runtime, as a backend of AMD's Lemonade Server and its OpenAI-compatible API. Boot a USB stick, and the machine serves local LLMs on its GPU and NPU. That is the whole setup.

Get started

1. Download the 1bit OS image from Hugging Face: 1bit-os-puppy.img.xz (about 2.2 GB compressed, 9,128,902,656 B written) with the SHA256SUMS beside it. The image is served from Hugging Face because a GitHub release asset is capped at 2 GiB; the releases page stays the home of the lemonade-onebit package, the release notes and the release SHA256SUMS, and links there for the image. The Hugging Face repository is private until v2026.44 is published, so until then that link answers 401; from publication day this step is the download.

2. Verify it, then write it to a USB stick (16 GB or more):

   sha256sum -c SHA256SUMS --ignore-missing
   xz -dc 1bit-os-puppy.img.xz | sudo dd of=/dev/sdX bs=4M conv=fsync status=progress

   balenaEtcher and Rufus (DD mode) write the .img.xz directly.

3. Boot the machine from the stick. Where our key is already enrolled, Secure Boot can stay on: the chain is a Microsoft-signed shim → our MOK-signed grub → our MOK-signed kernel → our signed modules. A machine that has never trusted our key refuses the kernel, and the certificates are not on the stick, so enroll both public ones from the box that signed the image first — or turn Secure Boot off for now; see docs/install.md.

4. Lemonade comes up with its catalog and web UI. Try it from the live session; to keep it, run 1bit-install <disk> --yes there and the same system is on the disk. The install options, what each one erases and the Secure Boot notes are docs/install.md; connecting a client — curl, the web app, the phone app, the OpenAI SDKs — is docs/clients.md.

1bit OS is a Puppy Linux appliance: ResolutePup64, built with woof-CE on the Ubuntu 26.04 "resolute" base, with Lemonade Server and this engine embedded in it, running our MOK-signed 7.4.0-rc1-1bit kernel rather than the distribution's. It boots headless by default, to a console of its own with no X, and a second boot entry, 1bit OS (desktop, JWM), starts Puppy's desktop on the same system. The userland is an ordinary one - bash, sudo and apt work - and everything that has to last (models, Lemonade's config and cache, SSH keys) lives on the 1BIT-DATA partition (os/README.md, docs/os-v3.md).

Already running Lemonade on Ubuntu? The lemonade-onebit package from the releases page adds the engine as a Lemonade backend; nothing else changes (docs/lemonade.md).

What it is: one model-agnostic binary for the GPU and the NPU

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.

Milestones

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)
2026-11-02	1bit OS becomes a Puppy Linux appliance: ResolutePup64 (woof-CE, Ubuntu 26.04 base) with our MOK-signed kernel, Lemonade and the engine embedded, headless by default with a desktop boot entry, apt and a persistent 1BIT-DATA; it boots on the board with Secure Boot on (docs/os-v3.md, os/README.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.

Documentation and measured results

• 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 serve and its options.
• docs/zero-copy.md: the copies left on the serving paths, what each one costs per token, and why it stays.
• Measured results: every number with how it was taken.
• Site repository: builds 1bit.gg from this repository.

License

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).

How to credit

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.

Thank you

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)	removed from the engine (owner 2026-11-01, out of the tree 2026-11-02): antirez's CUDA kernels through a HIP shim plus hipBLAS are not our HIP kernels, and --device ds4 now says it is gone. The fork 1bit-MONSTER/ds4 stays in the open as the reference for that work	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.