0.8.1

v0.8.1

v0.8.1 Integration Sprint — desktop wired end-to-end, zero mocks, Def…

0.8.0

v0.8.0

v0.8.0 Agentic Foundation — Tool Framework, ARIA Code, MCP Client, Se…

0.7.7

v0.7.7

v0.7.7 Wire Everything - ThinkingPipeline end-to-end, frontier retry,…

0.7.6

v0.7.6

v0.7.6 — Integration Sprint: SmartRouter, MoE routing, thinking toggl…

v0.7.5 — R&D + Documentation

What's New

ARIA Protocol v0.7.5 completes the R&D and documentation cycle following the v0.7.0 Smart Layer implementation.

Whitepaper v2

Complete protocol documentation (12 sections, 26 pages) covering the 5-layer architecture, Intelligence Stack, reputation system, competitive landscape, and roadmap.

PT-BitNet R&D (Research Result)

Comprehensive investigation of post-training ternarization for Qwen3-14B. Result: not viable — quality degrades 15-25 accuracy points at 14B scale, and no conversion pipeline exists for bitnet.cpp compatibility. Native 1-bit training remains the only practical path. Detailed report with tool analysis (onebitllms, BiLLM, PTQTP, ParetoQ).

KV-Cache NVMe Proof of Concept (aria/kv_cache.py)

Session persistence via llama-server's slot save/restore API. Save (~50ms) and restore (~43ms) conversation context between sessions. Foundation for v0.8.0's 500K+ token extended context.

Benchmark Report v2

Publication-quality analysis with 4 matplotlib charts (300 DPI): throughput comparison, native vs post-quantized (+42-50%), energy efficiency, and Zen 5 cross-generation speedup (+28-43%).

Accessibility Tree PoC (aria/accessibility.py)

Windows UI Automation reader that converts OS accessibility trees to structured text for LLM consumption. First building block for Computer Use. Cross-platform safe (graceful fallback on Linux/macOS).

Website Update

5-layer architecture diagram, Intelligence Stack section (MoE, Thinking Pipeline, Frontier API), v0.7.0 Smart Layer banner.

Stats

• 580 tests (531 → 580, +49 new tests)
• 3 new Python modules (kv_cache, accessibility, benchmark analysis)
• 4 commits in this release

Full Changelog

v0.7.0...v0.7.5

v0.7.0 — Smart Layer: Reputation + Intelligence

What's New

ARIA Protocol v0.7.0 introduces the Smart Layer — a complete intelligence stack for local 1-bit inference, with node reputation, multi-model routing, and structured reasoning.

Reputation System (aria/reputation.py)

• Node reliability scoring: uptime (40%), latency (20%), verification pass rate (40%)
• Efficiency bonus: energy-efficient nodes earn up to 2× score multiplier
• Anti-Sybil protection: subnet limiting, hardware fingerprinting, cluster detection
• Temporal decay with configurable half-life (default: 7 days)

Smart Router + MoE System Routing (aria/router.py)

• MoE dispatch: BitNet 0.7B (118 t/s) classifies queries and routes to the optimal expert model
• Routing matrix: task type × complexity → model selection (3B, 7B, or 10B)
• Confidence-based escalation: low-confidence local responses can escalate to frontier APIs
• Fully configurable routing policies

Frontier API Bridge (aria/api_bridge.py)

• 5 providers: Anthropic, OpenAI, Google, Mistral, xAI
• Local Fernet encryption for API key storage
• Per-provider usage tracking and cost estimation
• Injectable into SmartRouter as frontier_fn

Thinking Pipeline (aria/thinking.py)

• 3-phase Chain-of-Thought: THINK → RETRIEVE → GENERATE
• Phase 1 generates a structured reasoning plan (hidden from user)
• Phase 2 is a RAG hook ready for v0.8.0 knowledge engine
• Phase 3 produces the final response with enriched context
• Configurable: toggle per conversation, complexity threshold

Contribution Snapshot (aria/snapshot.py)

• Testnet contribution snapshots: uptime, inferences, energy per node
• Ranked leaderboard with Sybil exclusion
• JSON export/import for early adopter recognition

Documentation

• New: docs/reputation.md, docs/smart-router.md
• Updated: threat model (Sybil risk HIGH → MEDIUM), architecture (Layers 4-5), SECURITY.md

Stats

• 531 tests (338 → 531, +193 new tests)
• 5 new Python modules
• 7 commits in this release
• Zero breaking changes to existing APIs

Full Changelog

v0.6.0...v0.7.0

v0.6.0 - Testnet Alpha

`

ARIA Protocol v0.6.0 — Testnet Alpha

The largest release in ARIA's history. Transforms the protocol from a local inference engine into a fully functional distributed P2P network.

Highlights

Kademlia DHT — Decentralized peer discovery with O(log n) lookups, replacing static bootstrap lists. XOR distance routing, k-buckets, iterative FIND_NODE/STORE/FIND_VALUE.

NAT Traversal — STUN client (RFC 5389) for external IP discovery, UPnP IGD automatic port mapping, WebSocket relay fallback for symmetric NAT. Nodes behind home routers can now participate.

Ed25519 Authentication — Cryptographic node identity and message signing. Every message is signed with Ed25519, nonce + timestamp replay protection, keys persisted in ~/.aria/keys/.

50+ Node Simulator — In-process network simulator validated with 50 nodes: 100% shard discovery, 82.2% routing completeness, 1,892 connections, zero errors. Multi-process testnet launcher included.

Desktop P2P Visualization — New Network page with live WebSocket streaming of peers, Kademlia stats, NAT status, and shard registry. 12-language support.

Persistent Inference Server — llama-server.exe backend with model hot-swap (~1.5s), eliminating all subprocess crashes from v0.5.x.

Stats

• 122 commits since v0.5.5
• 82 files changed, +10,605 / -1,121 lines
• 338 tests (up from 196)
• 25 new files including 8 network submodules
• 0 errors in 50-node simulation

Network Architecture

aria/network/
├── core.py       — ARIANetwork (WebSocket P2P, message routing)
├── kademlia.py   — DHT routing table, k-buckets, XOR distance
├── nat.py        — STUN client, UPnP, WebSocket relay
├── auth.py       — Ed25519 signing, node identity
├── simulator.py  — 50+ node in-process simulator
├── pipeline.py   — Pipeline parallelism
├── tls.py        — TLS utilities
└── models.py     — PeerInfo, InferenceRequest

50-Node Simulation Results

Metric	Result
Nodes started	50/50 (100%)
WebSocket connections	1,892
Peers per node	avg 37.8 (min 22, max 50)
Kademlia routing completeness	82.2%
Shard discovery	100%
Peak RAM	372 MB (~7.4 MB/node)
Errors	0

Dependencies

• Added: pynacl>=1.5.0 (Ed25519 cryptography)
• Added (optional): miniupnpc>=2.2 (UPnP port mapping)

Full changelog

See CHANGELOG.md for complete details.

---

"The era of centralized AI infrastructure need not be permanent."

# ARIA Protocol v0.6.0 — Testnet Alpha

The largest release in ARIA's history. Transforms the protocol from a local inference engine into a fully functional distributed P2P network.

Highlights

Kademlia DHT — Decentralized peer discovery with O(log n) lookups, replacing static bootstrap lists. XOR distance routing, k-buckets, iterative FIND_NODE/STORE/FIND_VALUE.

NAT Traversal — STUN client (RFC 5389) for external IP discovery, UPnP IGD automatic port mapping, WebSocket relay fallback for symmetric NAT. Nodes behind home routers can now participate.

Ed25519 Authentication — Cryptographic node identity and message signing. Every message is signed with Ed25519, nonce + timestamp replay protection, keys persisted in ~/.aria/keys/.

50+ Node Simulator — In-process network simulator validated with 50 nodes: 100% shard discovery, 82.2% routing completeness, 1,892 connections, zero errors. Multi-process testnet launcher included.

Desktop P2P Visualization — New Network page with live WebSocket streaming of peers, Kademlia stats, NAT status, and shard registry. 12-language support.

Persistent Inference Server — llama-server.exe backend with model hot-swap (~1.5s), eliminating all subprocess crashes from v0.5.x.

Stats

• 122 commits since v0.5.5
• 82 files changed, +10,605 / -1,121 lines
• 338 tests (up from 196)
• 25 new files including 8 network submodules
• 0 errors in 50-node simulation

Network Architecture

aria/network/
├── core.py       — ARIANetwork (WebSocket P2P, message routing)
├── kademlia.py   — DHT routing table, k-buckets, XOR distance
├── nat.py        — STUN client, UPnP, WebSocket relay
├── auth.py       — Ed25519 signing, node identity
├── simulator.py  — 50+ node in-process simulator
├── pipeline.py   — Pipeline parallelism
├── tls.py        — TLS utilities
└── models.py     — PeerInfo, InferenceRequest

50-Node Simulation Results

Metric	Result
Nodes started	50/50 (100%)
WebSocket connections	1,892
Peers per node	avg 37.8 (min 22, max 50)
Kademlia routing completeness	82.2%
Shard discovery	100%
Peak RAM	372 MB (~7.4 MB/node)
Errors	0

Dependencies

• Added: pynacl>=1.5.0 (Ed25519 cryptography)
• Added (optional): miniupnpc>=2.2 (UPnP port mapping)

Full changelog

See [CHANGELOG.md](CHANGELOG.md) for complete details.

---

"The era of centralized AI infrastructure need not be permanent."`

ARIA Protocol v0.5.2 ÔÇö Multi-Backend Inference & Desktop Build Fixes

What's New in v0.5.2

Multi-Backend Inference

• Subprocess backend: Direct integration with bitnet.cpp via llama-cli for real 1-bit inference
• 3 backends available: native (ctypes), subprocess (llama-cli), simulation (testing)
• Auto-detection: ARIA automatically selects the best available backend

Real Benchmarks

Model	Params	Speed	Energy/token
BitNet-b1.58-large	0.7B	120.25 tok/s	~176 mJ
BitNet-b1.58-2B-4T	2.4B	36.62 tok/s	~869 mJ

Measured on AMD Ryzen 9 7845HX, 8 threads. Energy estimated via CPU-time × TDP.

Desktop App Fixes

• Version sync across all desktop configs (package.json, Cargo.toml, tauri.conf.json)
• Fixed macOS Apple Silicon (aarch64) build in CI/CD
• All 7 CI jobs passing (4 Tauri + 3 Electron)

Testing

• 176 tests passing (up from 102)
• New backend comparison benchmark suite

Downloads

See the assets below for your platform:

• Windows: .msi or .exe installer
• macOS Intel: .dmg disk image
• macOS Apple Silicon: .app bundle
• Linux: .AppImage, .deb, or .rpm package

Links

• Documentation
• Architecture
• Benchmark Results

Full Changelog: v0.5.1...v0.5.2

ARIA Protocol v0.5.1 - Desktop Build Fix

• Added homepage, author, description to package.json
• Added Linux maintainer to electron-builder.json
• Version bumped to 0.5.1

ARIA Protocol v0.5.0 - Desktop App

🖥️ Desktop App Release

Cross-platform desktop application with premium dark UI, available for Windows, macOS, and Linux.

Features

• Dashboard — Real-time stats, performance charts, activity feed
• Model Manager — Download and manage BitNet models from HuggingFace
• AI Chat — Local inference with conversation history and markdown support
• Energy Dashboard — Track savings, CO₂ avoided, achievements
• Settings — Full configuration with 12 languages

Supported Languages 🌍

English, Français, Español, Deutsch, Português, Italiano, 日本語, 한국어, 中文, Русский, العربية, हिन्दी

Tech Stack

• Tauri 2.0 — Lightweight (~15 MB)
• Electron — Maximum compatibility (~150 MB)
• React 18 + TypeScript + Tailwind CSS
• Framer Motion — Smooth animations

Downloads

Platform	Tauri (Recommended)	Electron
Windows	ARIA-Desktop_0.5.0_x64-setup.exe	ARIA-Desktop-Setup-0.5.0.exe
macOS Intel	ARIA-Desktop_0.5.0_x64.dmg	ARIA-Desktop-0.5.0.dmg
macOS Apple Silicon	ARIA-Desktop_0.5.0_aarch64.dmg	ARIA-Desktop-0.5.0-arm64.dmg
Linux	ARIA-Desktop_0.5.0_amd64.AppImage	ARIA-Desktop-0.5.0.AppImage

⚠️ Note: Pre-built binaries will be available once CI/CD builds complete. For now, build from source.

Build from Source

cd desktop
npm install
npm run tauri:build    # Tauri (recommended)
npm run electron:build # Electron alternative

Requirements

• Windows 10+ / macOS 11+ / Ubuntu 20.04+
• 4 GB RAM minimum
• 500 MB disk space

What's New Since v0.4.0

• Complete desktop application UI
• 12-language internationalization
• Tauri + Electron dual build system
• GitHub Actions CI/CD for automated releases
• Real-time performance monitoring
• Achievement system and gamification