Genesis Models
10 models ship at genesis, covering hardware from Raspberry Pi to server racks. All are GGUF format (Q4_K_M quantization), verified by SHA-256 digest from HuggingFace.
Compute nodes only download models they're configured to serve. A typical operator starts with 1-3 models matching their hardware tier.
Tier 1: Edge — 1-4 GB, CPU-only or 4GB VRAM
Raspberry Pi 5, smartphones, older laptops.
| Model | Params | Size | SHA-256 |
|---|---|---|---|
| Llama 3.2 1B Instruct | 1.2B | 1.32 GB | ba345c83...d9ae62a9 |
| Qwen3.5 4B | 4B | 2.74 GB | 00fe7986...69ef11a4 |
| Gemma 4 E2B Instruct | 2.3B active | 3.11 GB | ac0069eb...ee576845 |
Tier 2: Gamer — 4-8 GB, RTX 3060 or M3/M4 Mac
Entry-level GPUs and Apple Silicon laptops.
| Model | Params | Size | SHA-256 |
|---|---|---|---|
| Qwen2.5 7B Instruct | 7B | 4.68 GB | 65b8fcd9...ceaaa1423 |
| Llama 3.1 8B Instruct | 8B | 4.92 GB | 7b064f58...3033557c |
| Qwen3.5 9B | 9B | 5.68 GB | 03b74727...daf52b7e8 |
Tier 3: Prosumer — 16-24 GB, RTX 4090 or M-series Pro
High-end consumer GPUs and workstation Macs.
| Model | Params | Size | SHA-256 |
|---|---|---|---|
| Gemma 4 26B A4B (MoE) | 25.2B (3.8B active) | 16.87 GB | b8707e57...5a49fde |
| Qwen3.6 27B | 27B | 16.82 GB | 5ed60d0a...638392a0 |
| Gemma 4 31B (dense) | 30.7B | 18.32 GB | 3bf13fff...ea853474 |
Tier 4: Server — 48+ GB VRAM
Multi-GPU workstations and cloud instances.
| Model | Params | Size | SHA-256 |
|---|---|---|---|
| Qwen3.6 35B A3B (MoE) | 35B (3B active) | 22.1 GB | verify after download |
Adding models after genesis
Anyone can register a new model by submitting a RegisterModel transaction with:
- 10,000 ARK deposit (returned if model is valid)
- Model ID — canonical name (e.g.
meta-llama/Llama-3.1-8B-Instruct) - SHA-256 digest — of the GGUF artifact
- Size in bytes — for download verification
- Mirror URLs — where compute nodes can pull the file
- License — SPDX identifier
Compute nodes verify the SHA-256 after downloading — tampered files are rejected before loading into the inference engine.
Model sources
All genesis models are sourced from HuggingFace. Primary GGUF providers:
- unsloth — Qwen, Gemma quantizations with Dynamic 2.0
- bartowski — Llama, Qwen standard quantizations
- hugging-quants — Official HuggingFace quantizations