INTELLIGENCE.
UNBOUND.
The 1.58-bit 120B Large Language Model. Zero multiplications. 80% less compute. Pure reasoning.
Discover
The 1.58-bit 120B Large Language Model. Zero multiplications. 80% less compute. Pure reasoning.
BitNet architecture fundamentally rewrites the rules of artificial intelligence compute.
By quantizing weights to precisely -1, 0, and 1, we eliminate massive matrix multiplications. We rely strictly on simple additions, unlocking unprecedented speed.
Run massive 120B reasoning models directly on consumer hardware.
Less VRAM Required
Unprecedented energy efficiency without sacrificing reasoning. Extreme reduction in power cost.
Redefining production scale. A small GPU cluster handles real-time chat, while 70% of workloads (API/Batch) offload to massive, air-cooled CPU clusters using integer math. 90% fewer GPUs. Zero water cooling for CPUs.
The current trajectory of dense AI (like ChatGPT) is unsustainable. Nantiro solves this at the architectural level.
ChatGPT Monthly
Nantiro: Near Zero (Air Cooled)
ChatGPT Monthly
Nantiro: ~1M kWh (10x Less)
OpenAI Monthly Burn
Nantiro: 5-8x Cheaper
| Feature | Dense LLMs (OpenAI, etc.) | Nantiro (1.58-Bit) |
|---|---|---|
| Weight Representation | 16-Bit Floating Point (FP16) | Ternary {-1, 0, 1} |
| Core Mathematics | Complex Matrix Multiplication | Simple Integer Addition |
| Memory Footprint (120B) | ~240 GB VRAM | ~25 GB (9.6x Compression) |
| Hardware Requirements | Massive $30K+ GPU Clusters | Hybrid (Commodity CPUs + GPUs) |
| Compute Cost | Extremely High (Billions) | 80% Cheaper per Token |
| Environmental Impact | Millions of Liters of Water & Power | Air Cooled, Low Energy Burn |
| Model | Total Params | Active Params | Layers | Hardware |
|---|---|---|---|---|
| Phase 0 Tiny | 150 Million | 0.15B | 8 | 1 GPU/CPU |
| Nantiro 1B | 1.26 Billion | 1.26B | 24 | 8× H100 |
| Nantiro 7B | 6.72 Billion | 6.72B | 32 | 32× H100 |
| Nantiro 30B | 30.71 Billion | 30.71B | 64 | 128× H100 |
| Nantiro 120B (Phase 1) | 121.53 Billion | 121.53B | 96 | 1024× H100 |
| Nantiro 400B MoE | 396.94 Billion | 190.38B | 96 | 2048-4096× H100 |
| Nantiro 800B MoE | 802.52 Billion | 386.98B | 112 | 4096-8192× H100 |
| Nantiro 1T MoE | 1.005 Trillion | 399.87B | 112 | 8192+× H100 |
| Phase | Team Size | Network | Local NVMe |
|---|---|---|---|
| Phase 0A (1B) | 5-7 (ML, Data, Founder) | NVLink | 8-15TB |
| Phase 0B (7B) | 7-10 (+ Distributed, MLOps) | 200/400G IB | 30-60TB |
| Phase 0C (30B) | 10-14 (+ Safety, Evals) | 400G IB | 250-500TB |
| Phase 1 (120B) | 15-25 (24/7 Run Support) | 400/800G IB | 2-4PB |
| Phase 2 (400B MoE) | 30-45 (MoE, Kernel, Storage) | 800G IB | 4-8PB |
| Phase 3 (800B MoE) | 50-75 (Frontier-scale ops) | 800G+ IB | 8-16PB |
| Phase 4 (1T+ MoE) | 80-120 (Research Org Level) | 800G+/Next-gen | 16PB+ |
On-Premises Licensing (₹5L-₹20L/yr)
Serverless API (₹0.01/1k tokens)
Nantiro Chat Plus (₹199-₹499/mo)
Nantiro Edge Offline (₹999 once)
A: Direct GPU Provision
B: Direct Equity / SAFE
C: Cloud Credit Sponsorship