Aria Compute

huggingface.co/ariacompute · ariacompute.com · github.com/ariacompute · Shanghai, China

On-device AI, cloud-cooperative. Aria Compute builds an end-to-end quantized model and inference engine stack that brings production-grade LLMs, VLMs, and VLA models to phones, robots, wearables, and embedded devices — all running purely on-device CPU, with no network dependency.

Who We Are

Aria Compute is an AI infrastructure company founded in June 2026, focused on on-device multimodal foundation models and hybrid inference infrastructure. Our core offering is the Aria Compute platform: a unified toolchain for voice, vision, text, and embodied intelligence (VLA) models, built from the ground up for mobile and edge hardware.

What We Ship

Aria Engine

A from-scratch, pure-Rust inference runtime optimized for our quantized model bundles. Features include:

CPU-only execution — No GPU or cloud required
Blocked Hadamard domain multiplication (HDM) — Applies the same rotation tiles used during quantization to activations at runtime, enabling exact dequantization-free matrix multiplication
mmap-based weight loading — Direct memory-mapped IO with zero-copy weight access, minimizing startup latency and RAM overhead
Quantized KV cache — 4-bit per-entry cache reduces context memory ~4× vs FP16, keeping multi-turn conversation viable on low-RAM devices
Streaming token generation — Low-latency output suitable for real-time chat and tool-calling loops

Aria Quant Bundles

We quantize, package, and distribute production-ready bundles of popular open-source models. Every bundle is produced by our Python quantization toolkit using a common recipe stack:

Recipe	Strategy	Best For
q4 (`--bits 4`)	Uniform 4-bit, per-group codebooks (group_size=32)	Default baseline — smallest bundle
q8 (`--bits 8`)	Uniform 8-bit, per-group codebooks (group_size=32)	Near-lossless — maximum fidelity
q326_channel (`--bits 3.26 --codebook-share channel`)	Mixed precision (sensitive layers 4-bit, others ~3-bit), per-channel codebooks	Recommended generation-quality recipe — optimal quality-size trade-off

All recipes share the same pipeline: Hadamard rotation preprocessing followed by Lloyd-Max codebook quantization. The Hadamard transform rotates weight matrices to decorrelate columns, making them amenable to low-bit codebook compression with minimal reconstruction error. No calibration data is required — the method is data-free and deterministic given a random seed.

Each bundle ships as a lightweight directory: weight.bin (codebook-indices + per-group/per-channel codebooks) + config.json + tokenizer files. No GGUF, no safetensors — just Aria-native format.

Model Portfolio

We provide quantized bundles across 9 model families from the open-source community, each with q4 / q8 / q326_channel variants:

Family	Models	Domain	Key Partners
Qwen3	0.6B, 1.7B, 3.5-0.8B, 3.5-2B	Text generation, chat, tool-calling	Alibaba Cloud
Gemma	2-2B/9B, 3-270M/1B, 3n-E2B/E4B, 4-E2B/E4B	Text generation, chat, tool-calling	Google
LFM	LFM2 350M–8B-A1B, LFM2.5 350M–2.6B, VL 450M/1.6B	Text + vision, Liquid Neural Networks	Liquid AI
Nanbeige	4.2-3B	Chinese LLM	Nanbeige
Bonsai	1.7B, 27B	General text generation	Prism
Inkling	Small	General text generation	Thinking Machines
OpenVLA	7B	Vision-Language-Action (robotics)	Stanford / OpenVLA
OpenPI	Pi0-3B, Pi0.5-3B	Vision-Language-Action (robotics)	Physical Intelligence / LeRobot
LingBot	VLA-v2-6B	Vision-Language-Action (robotics)	Robbyant

Bench Evaluation

We maintain an automated cross-engine benchmark suite at bench/ that compares Aria-quantized models head-to-head against the same base models running in their officially recommended engines (llama.cpp GGUF, HuggingFace Transformers). Metrics span:

Capability: MMLU, GSM8K, C-Eval, HumanEval, Needle-in-Haystack, IFEval
Performance: TTFT, tokens/sec, peak memory, model size, power draw
Output quality: Token overlap, exact prefix fraction, logprob delta vs FP16

Design Philosophy

Local-first. All inference happens on-device. No data leaves the device. No network calls during inference.
Calibration-free quantization. Hadamard + Lloyd-Max requires no calibration dataset, no fine-tuning, no distillation — just the weight matrix.
One bundle, any CPU. ARM (Android, iOS, Linux) and x86_64 are first-class targets. Rust portability ensures the same bundle runs everywhere.
Progressive quality. Choose the recipe that fits your device: q4 for smallest footprint, q8 for near-lossless, q326_channel for the sweet spot in between.

Getting Started

HuggingFace: huggingface.co/ariacompute — Browse and download quantized model bundles
Website: ariacompute.com — Dashboard and documentation
GitHub: github.com/ariacompute — Engine, model quantization toolkit, and benchmark suite
Dashboard: ariacompute.com/dashboard/models — Authenticated model download

License

All Aria Compute quantized bundles inherit the license of their base model. See individual model cards for details. The Aria Engine and quantization toolkit are separately licensed.