Llama 3.2 3B (Standard)
The workhorse of the network. Optimized for extreme low-latency processing, this model is ideal for rapid classification, sentiment analysis, and straightforward data structuring.
For developers
Compute at a fraction of hyperscaler cost
A high-throughput distributed compute network powered by Apple Silicon Macs. REST API for async job processing and Server-Sent Event (SSE) streaming. Pay per token, no commitments, no quotas. EU-sovereign by design.
Utility tier notice
Utility Tier compute: data is protected in transit via standard TLS but is not processed in isolated hardware memory. Designed for non-sensitive public data, web scraping, and synthetic data generation. You are responsible for sanitizing all PII and PHI client-side before submission.
Async REST + SSE
stream long-running jobs
Pay per token
no commits, no quotas
EU-sovereign
data stays in-region
Zero egress fees
stream as much as you want
Batch API
high-throughput async jobs
Inference Compute API
sub-second cold start
Enterprise savings calculator
M tokens / mo
€26,777
Azure today
€9.33 / 1M tok
€13,447
AI Mining
€93.71 / 1M input tok
You save
€13,330/ month (50%)
1:20 Generative Envelope
Pricing is optimized for generative workloads (Chatbots & Agents) using our 1:20 Generative Envelope. For every 1 input token you buy, you get up to 20 output tokens included. You pay strictly for the prompt. The generation is on us.
Price comparison
What 1M Tokens Really Cost
Total price for 1 million input tokens plus 1 million output tokens.
$60.00
Claude
Anthropic
AI Mining 81% cheaper
$60.00
ChatGPT
OpenAI
AI Mining 81% cheaper
$60.00
Cursor
Anysphere
AI Mining 81% cheaper
$14.00
Gemini
AI Mining 17% cheaper
$11.60
AI Mining
us!
| Who | Model | 1M input | 1M output | Total |
|---|---|---|---|---|
| Claude (Anthropic) | Claude Fable 5.1 | $10.00 | $50.00 | $60.00 |
| ChatGPT (OpenAI) | GPT-6 Astra | $10.00 | $50.00 | $60.00 |
| Cursor (Anysphere) | Claude Fable 5 | $10.00 | $50.00 | $60.00 |
| Gemini (Google) | Gemini 3.1 Pro | $2.00 | $12.00 | $14.00 |
| AI Mining | Qwen 3.8 27B | $5.80 | $5.80 | $11.60 |
Prices from each company's public price list for their top model, checked October 5, 2026. Prices are checked automatically every week.
Models
Orchard AI Models: High-Performance AI Mining Batch
The Orchard network provides massive-scale, decentralized AI batch powered exclusively by our distributed Apple Silicon fleet. By shifting compute from centralized, industrial data centers to a global mesh of Mac hardware, we deliver frontier-level intelligence with zero water consumption, zero noise pollution, and a radically reduced carbon footprint.
Our model library is specifically optimized for Apple's Unified Memory Architecture (UMA), allowing us to offer large parameter models with unmatched cost-efficiency for batch extraction, structured data processing, and 3D asset generation.
I.
Text & Reasoning Models
Our text models are engineered to handle massive batch workloads—processing everything from localized JSON extraction to complex multi-document reasoning at unprecedented speeds.
Qwen 3.7 Max MoE (High Efficiency)
Our premier Mixture-of-Experts architecture. While it contains the vast intelligence of a 35-billion parameter foundation model, it activates only a fraction of those parameters per token. This provides enterprise-grade reasoning and logic capabilities at the processing speed—and price point—of a much smaller model.
Qwen 3.8 27B (Heavy Compute)
Built for the most demanding extraction and reasoning tasks. Thanks to our high-RAM network nodes (32GB+ to 352GB Exo clusters), this dense model runs natively in unified memory with massive context windows, eliminating the PCIe bandwidth bottlenecks that cripple traditional PC setups.
II.
Vision & Spatial Models
AI Mining's multi-modal capabilities allow buyers to process unstructured visual data, extracting precise text, bounding boxes, and spatial context from dense imagery.
Qwen 2.5-VL 3B
A lightning-fast vision model perfect for standard OCR tasks, receipt scanning, and basic image tagging.
Qwen 3-VL 8B
Our core production vision engine. It seamlessly balances high-resolution spatial understanding with rapid throughput, making it the standard choice for complex document ingestion (charts, graphs, and UI analysis).
Qwen 3-VL 32B (BETA)
Our flagship multi-modal powerhouse. Currently in beta for heavy network nodes, this model delivers frontier-level spatial intelligence and deep image comprehension for highly complex visual reasoning tasks.
III.
3D Asset Generation Models
Our decentralized compute grid powers the next generation of spatial intelligence and object creation, transforming 2D inputs into rich, production-ready 3D models.
Trellis 3D
A state-of-the-art 3D asset generation system. Trellis synthesizes high-quality 3D objects from a single text or image prompt in just a few seconds. Utilizing a novel Structured LATent (SLAT) representation, this model captures rich geometric shapes and surface textures. It uniquely decodes into multiple output formats simultaneously—including meshes (GLB), 3D Gaussian Splatting, and Radiance Fields (NeRF)—making it a highly versatile building block for game engines, AR/VR, and e-commerce.
IV. The Orchard Green Edge
A New Paradigm for AI Compute
Traditional AI hyperscalers rely on massive industrial data centers that extract a heavy toll on local environments. Orchard flips this model entirely. Our "data centers" are the Mac Minis and Mac Studios already sitting in living rooms, dorms, and offices around the world.
Zero Water Consumption
Traditional AI inference is incredibly water-intensive; a single training run for large models can consume 750 million liters of water for cooling alone. Standard data center cooling systems reject heat by evaporating roughly 1.8 gallons of water per ton-hour of cooling. Orchard uses exactly zero gallons of water for cooling. Our distributed Apple Silicon nodes rely entirely on ambient room temperature and hyper-efficient passive/active air cooling built into consumer hardware.
Ultra-Low Carbon Footprint
Because the GPU and CPU share a single physical memory pool, Apple Silicon eliminates the massive power drain of pushing data across traditional PCIe or NVLink buses. During sustained inference, an M4 Pro chip draws only 0.47 watts for the CPU/GPU package, with a total system power draw of just 8 to 12 watts. This delivers a 30x to 40x better energy efficiency per token compared to traditional data center GPUs, radically lowering the carbon emissions required to process your batch workloads.
Zero Noise Pollution
Hyperscale data centers generate relentless, industrial-grade noise pollution that disrupts local communities. Orchard's network is virtually silent. By utilizing consumer Apple hardware, our nodes execute heavy workloads while producing minimal thermal output and zero localized noise pollution, making our network completely invisible to the communities where our compute is generated.
AI MINING // Security & integrity
Hardened from packet to persistence.
Four layers of defensive architecture engineered specifically for distributed Apple Silicon nodes.
01 / 04
Runtime Type Integrity
The logic
Eliminate “garbage in, garbage out” at the edge.
The job
Strict schema validation on every ingress point. Runtime type-safety drops malicious payloads — like strings injected into numerical fields — before they ever touch the logic layer. This architecture is what maintained stability during our 1B-entry stress test.
02 / 04
Continuous SAST & Vulnerability Scanning
The logic
Mitigate supply-chain attacks and human error.
The job
Our CI/CD pipeline integrates Static Analysis Security Testing to audit the codebase in real-time. It proactively hunts for leaked secrets, insecure dependencies, and logical hidden holes, keeping the stack enterprise-ready and resilient against automated exploits.
03 / 04
Heuristic Threat Detection
The logic
Signature-based AV is too high-latency for real-time streams.
The job
Rigid application-layer verification designed for telemetry packets. The gateway enforces strict schema matching and validation constraints on all incoming payloads, dropping malformed inputs or irregular parameters before they can touch downstream processing node loops.
04 / 04
End to End Encryption
The logic
Defending distributed communication vectors.
The job
Multi-layered transit cryptography. All data payloads routed across the orchestrator plane to processing edge nodes are encapsulated within secure TLS 1.3 cryptographic tunnels, protecting data-in-transit from structural interception risks.
AI MINING // Infrastructure
The substrate beneath the swarm.
A scalable backbone purpose-built to coordinate, visualize, and persist a global AI Mining fleet.
01 / 04
Scalable Relational Backbone
The logic
High-concurrency persistence with deterministic tracking.
The job
A battle-tested relational database featuring native Row-Level Security (RLS) and continuous, timestamped system audit trails. It serves as the authoritative cloud coordinator for the AI Mining network, managing global authentication and active worker state at scale.
02 / 04
High-Performance Geospatial Visualization
The logic
Precision rendering without the overhead of heavy GIS suites.
The job
SVG-based vector grid rendering powers our Command Center — a lightweight, government-grade geospatial interface for the Copenhagen Genesis Node with sub-millisecond interaction speeds.
03 / 04
Optimistic State & Stream Visualization
The logic
Synchronous UI/UX for asynchronous hardware.
The job
Advanced cache management and optimistic UI updates keep the dashboard in sync with Mac Mini M4 hardware. Our Live Throughput telemetry uses high-frequency polling to visualize network health without a single page refresh.
04 / 04
Zero-Downtime Edge Deployment
The logic
Ship without breaking the swarm.
The job
Atomic edge rollouts with instant global propagation and one-click rollback. Versioned worker contracts let us push new dispatcher logic to thousands of Mac Mini M4 nodes without dropping a single in-flight job — the fleet keeps earning while we deploy.
Powered by Apple M-series silicon
Security carved into Apple Silicon, not bolted on.
Every node in our fleet is a Mac running inside Apple's own walled garden — the same hardware-rooted trust model that protects two billion consumer devices, now coordinating sovereign AI compute.
Keychain storage
100%
isolated from config
System integrity
Native
macOS hardware trust
Hyperscaler GPU CVEs
0
applicable vs. ANE on M4
01 / 04
Keychain-Backed Credentials
The logic
Native macOS credential storage isolated from configuration files.
The job
Nodes leverage native macOS storage frameworks, using the system Keychain to isolate and persist host authentication credentials. Network identity tokens never sit in plaintext configuration files. Execution sandboxing relies on standard operating-system process segmentation and memory access controls.
02 / 04
Hardware-Verified Boot Chain
The logic
Trust starts before the kernel does.
The job
Apple Silicon enforces a cryptographically verified boot process from the hardware Boot ROM up through the signed system kernel. Host edge environments inherit this secure execution sequence, providing a tamper-resistant system runtime baseline before processing any jobs.
03 / 04
Isolated Process Space Execution
The logic
The same separation perimeter used to isolate system runtimes.
The job
The inference daemon executes inside an independent, local process memory space bound strictly to a non-public loopback interface (127.0.0.1). By restricting local port exposure, buyer processing workloads are segmented away from standard remote execution entryways.
04 / 04
On-Device Neural Engine
The logic
Inference never leaves the silicon.
The job
All token generation runs on the Apple Neural Engine and unified memory of the M4 — no PCIe bus, no external GPU, no hyperscaler tenancy. The attack surface of a traditional cloud GPU (shared VRAM, side-channel leakage, hypervisor escapes) simply does not exist on our fleet.
Net result
A typical hyperscaler GPU instance trusts the hypervisor, the firmware, the driver stack, and the shared tenancy layer — four perimeters you don't control. Our Apple Silicon fleet collapses that chain into localized, hardware-rooted execution perimeters. The result is a measurably tighter containment boundary and a significantly smaller attack surface for your live inference workloads.
Optimizing data density & sharding
The Performance Principle: High Density, Low Overhead
To achieve maximum throughput on the AI Mining grid, we recommend optimizing your data density before submission. While our architecture supports massive concurrency, the speed of your upload and subsequent processing is heavily influenced by Request Volume versus Data Volume.
Every document submitted to the grid requires a cryptographic handshake, database indexing, and a security scan. If you submit 100,000 documents of 1KB each, you incur the "latency tax" of 100,000 individual operations.
For optimal performance:
- Target Document Size: Aim for 1MB to 10MB per shard.
- Consolidate Records: Instead of submitting one document per data row or sentence, bundle your data into larger JSON objects or files.
- Node-Side Efficiency: Our decentralized nodes (utilizing Apple Silicon M-series hardware) are optimized for high-speed streaming. They can "unpack" and process a 10MB document much more efficiently than they can manage thousands of 1KB pings.
| Feature | Fragmented data (sub-optimal) | High-density data (optimized) |
|---|---|---|
| Document Size | 1KB - 50KB | 1MB - 10MB |
| Network Overhead | High (Per-request latency) | Low (Sustained throughput) |
| Processing Speed | Limited by I/O handshakes | Limited only by Node CPU |
| Effective Cost | Standard | Maximum ROI |