AMD Strix Halo · Ryzen AI Max+ 395

⚡ One binary to rule them all

Run AI models on
your own hardware.
Fast.

One C++ binary. Zero Python. MIT licensed.

Including on a chip nothing else can touch: AMD's own runtime for the Strix Halo NPU is closed-source, so we reverse-engineered it from scratch. Also runs on AMD / NVIDIA / Apple GPUs, any Vulkan GPU, and plain CPU. It's an inference engine, not a chat app — bring your own frontend. 47 models run out of the box, 135M to 74B parameters.

View on GitHub →Get Started →

1bit.systemsSTRIX HALO

NPU MODELS

ARCHITECTURES

BACKENDS

Top decodetok/s · 2026-08-01

SmolLM2-135M

662

Qwen3-0.6B

373

Qwen2.5-VL-3B

100

BlackMamba-1.5B

79.4

C++23 · MITzero PythonNPU unlocked1.5 MB binary

Origin Story

22 proprietary libraries → one 1.5 MB binary

We reverse-engineered the whole stack in 4 days and replaced it with open C++ — turning 22 proprietary libraries into one 1.5 MB open-source binary.

AMD shipped the Ryzen AI Max+ 395 with a 50 TOPS NPU locked behind a proprietary driver stack. No documentation. No open SDK. No way to run your own models.

1,800+ hours of reverse engineering, every crash documented in real time. The result: a fully open inference engine that runs on the bare metal.

Before

22 proprietary libraries

Closed-source driver stack, no docs

After

1.5 MB open-source binary

Pure C++, no Python in the hot path

Read the full journey →

.1bp Format

Our models, our format

Every model is packed into a single .1bp file — weights, tokenizer, and config. No GGUF conversion. No Python. Just load and run.

GGML-Vulkan135M

SmolLM2-135M

662tok/s

GGML-Vulkan0.6B

Qwen3-0.6B

373tok/s

GGML-Vulkan3B

Qwen2.5-VL-3B

100tok/s

GGML-Vulkan4B

Qwen3.5-4B

65tok/s

Mamba1 HIP1.5B

BlackMamba-1.5B

79.4tok/s

GGML-Vulkan8B

DeepSeek-R1-Distill-Llama-8B

44tok/s

Voice Assistant

JARVIS — fully local voice AI

No cloud. No Python in the hot path. Every stage runs on the engine — from microphone to cloned-voice speaker output.

🎙️

mic

🔊

VAD

📝

STT

Whisper

📡

router

🧠

LLM

📢

TTS

codec

🎵

cloned voice

🔈

speaker

JARVIS Docs →

Model Ecosystem

19 architectures. 47 models. 17 families.

From 135M to 8B+ parameters — text, vision, code, and speech models packed into .1bp format.

ZyphraQwenLlamaMistralGemmaPhiFalconOLMoGraniteSmolLMDeepSeekGPT-OSSLagunaKimiBitNet/BonsaiWhisper

Full Catalog

16 models in .1bp format

Model	Family	Params	Arch	Backend	Gen tok/s
SmolLM2-135M	SmolLM	135M	LLaMA	GGML-Vulkan	662
Qwen3-0.6B	Qwen	0.6B	Qwen2	GGML-Vulkan	373
Qwen2.5-VL-3B	Qwen	3B	Qwen2VL	GGML-Vulkan	100
Qwen3.5-4B	Qwen	4B	Qwen2	GGML-Vulkan	65
BlackMamba-1.5B	Zyphra	1.5B	Mamba1	Mamba1 HIP	79.4
DeepSeek-R1-Distill-Llama-8B	DeepSeek	8B	LLaMA	GGML-Vulkan	44
Phi-3-mini-4k	Phi	3.8B	Phi3	GGML-Vulkan	—
Gemma-2-2B	Gemma	2B	Gemma2	GGML-Vulkan	—
Llama-3.2-1B	Llama	1B	LLaMA	GGML-Vulkan	—
Mistral-7B-v0.3	Mistral	7B	LLaMA	GGML-Vulkan	—
Falcon-1B	Falcon	1B	Falcon	GGML-Vulkan	—
OLMo-1B	OLMo	1B	OLMo	GGML-Vulkan	—
Granite-3B-Code	Granite	3B	LLaMA	GGML-Vulkan	—
Kimi-VL-A3B	Kimi	3B	MoonViT	GGML-Vulkan	—
Whisper-tiny	Whisper	39M	Whisper	GGML-Vulkan	—
Whisper-base	Whisper	74M	Whisper	GGML-Vulkan	—

Performance

662 tok/s peak generation

Measured 2026-08-01 on AMD Ryzen AI MAX+ 395, Radeon 8060S, 32 GB UMA.

43.2 TFLOPS INT8 prefill (WMMA)

SmolLM2-135M🏆

GGML-Vulkan662tok/s

Qwen3-0.6B

GGML-Vulkan373tok/s

Qwen2.5-VL-3B

GGML-Vulkan100tok/s

BlackMamba-1.5B

Mamba1 HIP79.4tok/s

Qwen3.5-4B

GGML-Vulkan65tok/s

DeepSeek-R1-Distill-Llama-8B

GGML-Vulkan44tok/s

Hardware Backends

5 primary backends + CPU & Metal

AMD NPU + GPU, NVIDIA CUDA, Apple Silicon, any Vulkan GPU, and x86 CPU. One binary targets them all.

NPUNPU

AMD XDNA 2 (Strix Halo)

50 TOPS, proprietary stack fully reverse-engineered

HIPGPU

AMD ROCm GPU

Radeon 8060S, 32 CUs

ZINCGPU

Any Vulkan 1.2+ GPU

SPIR-V compute shaders, multi-arch

GGML-VulkanGPU

Any Vulkan 1.2+ GPU

Broad compatibility, peak throughput

CUDAMULTI

NVIDIA GPU (sm_70+)

RTX 20-series and newer

MetalGPU

Apple Silicon

M1/M2/M3/M4 series

CPUMULTI

x86 OpenMP fallback

Runs anywhere, no GPU needed

Quick Start

Build from source

No installer yet — today it's build-from-source. CMake, a C++ compiler, and you're running in under a minute.

# Clone and build
git clone https://github.com/1bit-systems/1bit-systems
cd 1bit-systems && cmake -B build && cmake --build build
# Run inference
./build/1bit zaya -m model.1bp -p "Hello world"

No installer yet — build from source. See the Installation Guide →

FAQ

Run AI models onyour own hardware.Fast.

22 proprietary libraries → one 1.5 MB binary

Our models, our format

JARVIS — fully local voice AI

19 architectures. 47 models. 17 families.

Full Catalog

662 tok/s peak generation

5 primary backends + CPU & Metal

Build from source

Frequently asked questions

Run AI models on
your own hardware.
Fast.