Arc

Arc is a purpose-built, EVM-compatible Layer-1 blockchain advancing the frontier of stablecoin finance and tokenization. It features USDC as native gas, deterministic settlement finality, opt-in privacy, and a stable transaction fee architecture. Optimized for stablecoin-native use cases, such as global payments, FX, and capital markets, Arc serves as a foundational settlement layer for programmable money on the internet.

General
Company	Circle Internet Group
Website	Arc Network
Documentation	https://www.arc.network/litepaper
Type	Layer-1 Blockchain
Launch	Public Testnet Fall 2025, Mainnet Beta 2026
Consensus	Malachite (Byzantine Fault Tolerant)
Compatibility	EVM-Compatible

Core Architecture

Malachite Consensus Engine

High-performance Byzantine Fault Tolerant (BFT) consensus
Built on Tendermint algorithm with Rust implementation
Sub-second deterministic finality (under 350ms with 20 validators)
Performance up to 10,000 TPS with 4 validators
No probabilistic confirmations or chain reorganizations

USDC as Native Gas Token

Transaction fees paid directly in USDC stablecoin
Predictable, dollar-denominated costs for enterprises
Eliminates volatile crypto token exposure for gas payments
Enhanced EIP-1559 fee mechanism with weighted moving averages
Stable transaction fee architecture optimized for business use

EVM Compatibility

Full Ethereum Virtual Machine compatibility
Seamless migration of existing Ethereum applications
Support for existing developer tools and frameworks
Native integration with Solidity smart contracts

Key Features

Deterministic Settlement Finality

Transactions achieve irreversible finality in under 1 second
No risk of chain reorganizations or reversals
Guaranteed final settlement for enterprise applications
Superior to probabilistic finality models

Built-in FX Engine

Institutional-grade Request-for-Quote (RFQ) system
24/7 on-chain foreign exchange and settlement
Payment-versus-Payment (PvP) atomic swaps
Support for multiple stablecoin pairs (USDC, EURC, etc.)
Perpetual futures markets for stablecoin trading

Opt-in Privacy Controls

Confidential transfers hiding transaction amounts
Addresses remain visible for compliance
View keys for selective disclosure to regulators/auditors
Trusted Execution Environment (TEE) implementation
Future support for MPC, FHE, and Zero-Knowledge proofs

MEV Mitigation

Classification of constructive vs. harmful MEV
Encrypted mempools to prevent front-running
Batch transaction processing
Multi-proposer mechanisms for fair ordering
Protection against sandwich attacks and value extraction

Enterprise Integration

Circle Platform Integration

Native support for Circle Payments Network (CPN)
Full integration with USDC, EURC, and USYC tokens
Circle Mint, Wallets, and Contracts compatibility
Cross-Chain Transfer Protocol (CCTP) support
Gateway interoperability service integration

Real-World Asset Tokenization

Support for tokenized equities, bonds, and securities
Private credit and institutional fund tokenization
Partnership with licensed asset issuers and custodians
Regulated real-world asset (RWA) framework
Compliance-ready tokenization infrastructure

Institutional Validator Network

Permissioned Proof-of-Authority consensus initially
Hand-selected institutional validators
Geographic distribution of validator nodes
High compliance and regulatory standards
Future transition to broader validator participation

Use Cases

Global Payments & Remittances

Cross-border payments with instant settlement
Reduced intermediary costs and complexity
24/7 operation independent of banking hours
Multi-currency stablecoin support

Capital Markets & Trading

Institutional trading settlement
Tokenized securities and derivatives
Automated delivery-versus-payment (DvP)
Real-time collateral management

Foreign Exchange

On-chain FX trading and settlement
Stablecoin pair perpetual futures
Institutional-grade price discovery
Automated currency conversion

Treasury Management

AI-powered treasury optimization
Yield-bearing stablecoin integration (USYC)
Programmable finance workflows
Automated compliance reporting

Technical Specifications

Throughput: 3,000-10,000 TPS depending on validator count
Finality: Sub-second (100-350ms) deterministic settlement
Consensus: Malachite BFT based on Tendermint
Virtual Machine: Ethereum Virtual Machine (EVM) compatible
Gas Token: USDC native gas payments
Privacy: Opt-in confidential transfers with compliance features
Interoperability: Cross-chain bridges and CCTP integration

Development Timeline

Private Testnet: August 2025 (launched)
Public Testnet: Fall 2025
Mainnet Beta: 2026
Full Production: TBD with community and regulatory readiness

Arc represents Circle's vision for a stablecoin-native financial infrastructure that bridges traditional finance with programmable blockchain technology, designed specifically for enterprise adoption and regulatory compliance.

Dispatch

Most routing agents spend tokens deciding how to spend tokens. Ours doesn't: routing is a local forward pass that never touches the API, so every routing decision costs exactly zero tokens under Track 1 scoring. At the core is a fine-tuned DistilBERT categorizer (ONNX INT8, 67 MB, ~11 ms per decision) that classifies each task into one of the 8 hackathon categories. A measured policy then maps each category to the cheapest engine that can actually answer it, walking a local-first cascade: deterministic solvers compute exact answers for math (0 tokens), a local Qwen2.5-3B handles the factual lane (0 tokens, 8.4 s/request on the judge's AMD Zen 4 hardware), and only what remains escalates to Kimi on Fireworks — the single billed step. Why fine-tune instead of heuristics? Regex hit 114/116 on validation, but only by matching verbatim cue words — on a cue-free paraphrase slice it collapsed to 32%, while DistilBERT held 100% (116/116 overall). Low-confidence predictions on token-saving lanes fail safe to the strong model, and generated code is verified by sandboxed test execution before it's trusted. The policy isn't hand-tuned either: an exhaustive per-category sweep against a pre-labeled answer cache picks the Pareto winner on accuracy vs. tokens, with zero new API calls. Official 10-task validation: 10/10 accuracy at 2,519 tokens — 26% under always-Kimi (3,415) and 35% under always-cheap (3,851; the "cheap" model is priced lower per token but ~1.7× more verbose, so it loses on both axes). Shipped as a Docker container meeting the full judging contract at 2 vCPU / 4 GiB.

OmniMesh: The Phone That Calls for Help Itself

When a disaster hits, the first thing that goes down is the ability to ask for help. Cell towers fail, the internet disappears, and the people who need finding fastest are usually unconscious, trapped, or unable to reach a phone screen. We built OmniMesh to close that gap for victims, responders, and command teams at once. An Android phone left face-down and still after a hard impact notices on its own. On-device AI fusion runs continuously: injury classification, an acoustics model tuned for disaster sounds, a bidirectional LSTM reading motion for collapse signatures, multimodal assessment, fused into a confidence-scored decision in under 3 seconds. Detect collapse plus stillness, and it auto-broadcasts a RED packet with GPS. No button, no app opened. Phones form a decentralized Bluetooth mesh: RED-first routing, store-and-forward, hop-by-hop relay, mesh walkie-talkie, holding with zero towers or wifi. Every packet reaches all three roles. The victim gets voice-guided first aid from an AI companion. The responder sees exactly where to go, ranked by real urgency, on a live map. Command sees the incident unfold in real time: casualty estimates, zone assignments, resource recommendations. Underneath sits a multi-agent AI backend on real AMD hardware two ways. Gemma answers locally on our own AMD GPU through ROCm, working offline. Online, a cloud pass through Fireworks AI, also AMD infrastructure, reconciles the answer. A vision agent reads structural damage from photos. A matching agent reunites missing people with whoever found them. Every answer carries a confidence score, and when unsure, it flags the case for a person instead of guessing. Nothing here has a single point of failure. Lose the cloud, the phone still triages with plain medical rules. Lose everyone's internet, two phones five meters apart still find each other and keep the message moving. GitHub: https://github.com/Adya6714/OmniMesh Live demo: https://omnimesh-command.web.app

Hybrid Token Efficient Routing Agent

This project is a routing agent that decides, per task, whether a small local model can answer correctly or whether the query needs to be escalated to a larger model hosted on Fireworks AI. The core idea: a local Qwen model, run inside the container, attempts every task first. Alongside its answer, it self-reports a confidence score. If that score clears a threshold, the local answer is used and the task costs zero tokens toward the hackathon's scoring, since local inference isn't billed. If confidence falls below the threshold, the task is escalated to a Fireworks-hosted model via FIREWORKS_BASE_URL, using only models from the harness-provided ALLOWED_MODELS list. The agent is built to handle all eight task categories specified for Track 1: factual knowledge, mathematical reasoning, sentiment classification, text summarization, named entity recognition, code debugging, logical/deductive reasoning, and code generation. It reads tasks from /input/tasks.json and writes results to /output/results.json in the required schema, and reads all Fireworks credentials and model IDs from environment variables at runtime rather than hardcoding them, per the harness rules.

dkdsja

AMD Developer Hackathon: ACT II Three tracks. Real AMD hardware. Pick your challenge. Whether you're shipping your first AI agent or building your next startup, ACT II has a track for you. Build on AMD Developer Cloud, ROCm, and Fireworks AI API credits. All submissions must be containerized. Track 1 Hybrid Token-Efficient Routing Agent ⭐ Beginner · AI Agent Track Track 2 Video Captioning 🎬 Beginner · Prompt and have fun Track 3 Unicorn Track 🦄 All levels · Build your startup Track 1 ⭐ Beginner Friendly · AI Agent Track Hybrid Token-Efficient Routing Agent Build an AI agent that gets the job done using the least tokens possible. Tasks are revealed at kickoff. Your agent must complete each one autonomously by deciding in real time whether to use a local model or call a remote model via Fireworks AI credits. The goal: pick the cheapest option every time, without falling below the accuracy threshold. Every submission is scored on a standardized environment. You can develop and test on any hardware, but final scoring runs on this standardized environment only. Local models must therefore be sized to run within these constraints, so routing intelligence wins, not raw compute power. 💡 All models and tokens used locally count as zero toward the final score. We recommend running a local eval step to check your output quality before submitting. Want to fine-tune your router? Go for it. Prompt-based and fine-tuned approaches are scored exactly the same way: token count and output accuracy. Models to be used will be revealed on launch day. 💡 Build Ideas Model Router / Cost Optimizer A routing layer that reads each query and instantly picks the cheapest, best-suited model from the available endpoints. Level Beginner Judging Token count and output accuracy Compute Local model + Fireworks AI API

Recoverflow-Cross-border AR on autopilot.

Problem. A Taipei SME stuck in Net 60 with a US buyer has no good options: Atradius takes 8 months and 30%, a US attorney costs more than the outstanding, and QuickBooks emails sound like Google Translate. Existing tools assume your CFO speaks American English in the buyer's timezone — false on the Taiwan→US corridor. Solution. Recoverflow is a 9-agent dunning system on Band (8 production + AAA Specialist via Day-65 dynamic peer discovery): Pre-flight 3-path routing (in_spot / lite / attorney_recommended), Investigator pattern-tagging replies, Diplomat cadence-aware emails, Tone Coach (Claude) blocking hostile + FDCPA-non-compliant language, Escalator on Day 65, Voice Agent via ElevenLabs ConvAI + Twilio, Concierge as Slack HITL choke point, Payment Agent settling Circle USDC on real ARC-TESTNET. Outstanding, not gross (D-038). A $47,300 gross invoice the lawyer rejects as "out of sweet spot" is $23,650 outstanding after a 50% deposit — exactly where we operate. Wired through 5 modules across 4 demo Beats including the demand letter. Tech. Python 3.12, Band SDK, Claude Sonnet 4.5 / Haiku 4.5, ElevenLabs ConvAI, Twilio, Circle Programmable Wallets on real ARC-TESTNET (5 verified on-chain tx), Slack HITL, Featherless, Chroma RAG (31 chunks via BGE), pytest 482-test suite. Demo highlights. - Beat 4 — Pre-flight 3-path routing on outstanding balance. - Beat 8 — Welfare 988 SOP: set_anomaly_halt(case_id, "welfare") fires BEFORE Concierge page (life before debt, 3 unit tests). - Beat 9-11 — 5 real Circle USDC tx settled in data/audit_trail.jsonl (475dcb1c…, c6b4fe6b…, 36e503b2…, 2bcfdc5b…, 496dc964…). Track 3 fit. FDCPA language gating, paylink-only demand letters (no AI-hardcoded SWIFT/IBAN), HITL on every outbound, deterministic checkout URLs, 988 priority over debt, ConvAI 16-reason escalation enum, full audit_trail.jsonl chain of custody.