FIELD NOTES
SUPPLY·HARDWARE·Aug 23, 2026

Powering Decentralized AI: The DePIN.as Model Catalog & Hardware Matrix

DePIN.as connects high-performance Apple Silicon nodes into a global decentralized compute mesh. By combining Data Parallelism over high-speed networks with Model Parallelism over local 80 Gbps Thunderbolt bridges, DePIN.as delivers fast, cost-effective inference for Text and Vision workloads.

Page 1: The Complete Model & Hardware Matrix

Every job dispatched across the network is dynamic-RAM gated at ingest. Our automated claim engine routes tasks exclusively to nodes and clustered racks that meet exact unified memory floors.

ModalityModel KeyDisplay NameTierHardware Floor Requirement
Textllama-3.2-3bLlama 3.2 3BStandard8 GB+ Unified Memory
Textqwen-3.7-maxQwen 3.7 Max MoEPremium ("Big Rig")32 GB+ Unified Memory
TextQwen 3.8 27BQwen 3.8 27BPremium / Rack Fleet32 GB+ (≤ 32,768 context) · 48 GB+ (> 32,768 context)
Visionqwen2.5-vl-3bQwen 2.5 VL 3BStandard8 GB+ Unified Memory
Visionqwen3-vl-8bQwen 3 VL 8BMedium16 GB+ Unified Memory
Visionqwen3-vl-32bQwen 3 VL 32BPremium ("Big Rig")32 GB+ Unified Memory

Page 2: Deep Dive & Tiered Workload Routing

Standard & Medium Tiers (8 GB – 16 GB Unified Memory)

Designed for high-throughput, low-latency execution across standalone Apple Silicon hardware (e.g., M-series Mac minis and MacBook Airs):

  • Llama 3.2 3B & Qwen 2.5 VL 3B: Ultra-fast baseline models optimized for real-time document extraction, structured data generation, and rapid image parsing with sub-20 ms per-token latencies.
  • Qwen 3 VL 8B: Medium-tier vision engine servicing high-density image reasoning tasks while maintaining a lightweight 16 GB memory footprint.

Big Rig & Thunderbolt Rack Tiers (32 GB – 48 GB+ Unified Memory)

For heavy reasoning, complex code refactoring, and enterprise vision tasks, high-parameter models require aggregated compute:

  • Qwen 3.7 Max MoE & Qwen 3 VL 32B: Routed to standalone high-memory hosts (M-Max/Ultra chips with 32 GB+ RAM) to deliver frontier-class visual and mixture-of-experts logic.
  • Qwen 3.8 27B (The Developer Engine): Our flagship text model featuring dynamic context-aware memory routing:
    • Standard Context (≤ 32k tokens): Unlocked by single 32 GB machines or 2× 16 GB Thunderbolt 5 racks.
    • Extended Context (> 32k tokens): Automatically escalates to 48 GB+ memory pools to preserve zero-OOM execution stability.

Getting Started

  • For Buyers: Submit batch jobs targeting any active model key using our standard REST API endpoints.
  • For Node Operators: Earn premium compute rewards by connecting two or more M-series Macs via Thunderbolt to form 32 GB+ aggregate compute racks.