Aria Compute

ariacompute.com · github.com/ariacompute · huggingface.co/ariacompute

On-device AI, cloud-cooperative. Aria Compute builds hybrid inference infrastructure and foundation models for phones, robots, wearables, and edge devices.

Who We Are

Aria Compute is an AI infrastructure company founded in June 2026, focused on on-device multimodal foundation models and hybrid inference. On this Hub org we publish product checkpoints that run locally.

What We Ship

AFM-D — System 1 typed decisions

AFM-D answers typed questions over a text or JSON state and returns a distribution over the answers you supply. It is not open-ended chat and not TypeSafe Jev.

• Context max_len=1024 with shared head_max_len=512 for question + options.
• Decision temperature fixed at 1.0; confidence temperatures fitted per task / option-count via ECE.
• Encoder checkpoint: Laya layout (model.safetensors, rl_agent_config.json, tokenizer/, encoder/).
• Decoder checkpoint: PEFT adapter (adapter_*.safetensors, adapter_config.json, dd_config.json, tokenizer/).
• No GGUF export and no autoregressive chat path on either track.

Product gate: JevBench. Live board: Benchmark Heaven / JevBench Space.

Aria Engine

Separately, Aria ships a pure-Rust on-device inference runtime.

Design Philosophy

• Local-first. Inference stays on-device; no network required at decision time.
• Typed answers, not chat. AFM-D scores the options you declare — Choice, Score, or Noul.
• Two tracks, one product. Encoder for fast ModernBERT decisions; Decoder for SemIf + MiniCPM LoRA / System One HTTP.
• Honest gates. Release diagnostics (agreement / ECE / Brier) plus JevBench.

Getting Started

from huggingface_hub import snapshot_download

enc = snapshot_download("ariacompute/afm-de")
dec = snapshot_download("ariacompute/afm-dd")
# Load with the AFM-D package: --checkpoint <path>

License

Each model inherits the license of its base stack and training data. Aria Engine and tooling may be licensed separately.