Tinybox: Offline AI Device with 120B Parameters — George Hotz's Tinygrad Ships Red V2
Tinybox: Offline AI Device with 120B Parameters
George Hotz and the tinygrad team have announced that the Tinybox is now shipping — a purpose-built AI inference computer designed for running large language models entirely offline.
Available Models
Tinybox Red V2 — In stock, $12,000:
- 4x AMD Radeon RX 9070 XT GPUs
- 64 GB GPU RAM total
- 778 TFLOPS (FP16)
- 128 GB system RAM, 2 TB NVMe
- Runs 120B+ parameter models offline
Tinybox Green V2 (Blackwell) — In stock, $65,000:
- 4x NVIDIA RTX PRO 6000 Blackwell GPUs
- 384 GB GPU RAM total
- 3086 TFLOPS (FP16)
- 192 GB system RAM, 5 TB RAID
- Full fabric interconnect, PCIe 5.0
Exabox — Coming 2027, ~$10M:
- 720x AMD RDNA5 AT0 XL GPUs
- 25,920 GB (25 TB) GPU RAM
- ~1 EXAFLOP performance
- 20,000 lb enclosure requiring a concrete slab
Why It Matters
The Tinybox represents a growing movement toward personal, offline AI inference hardware:
- Privacy-first AI. No data leaves the machine. Enterprise and privacy-sensitive users can run powerful models without cloud dependencies.
- Democratization of AI compute. At $12,000 for the Red V2, running 120B parameter models is becoming accessible to research labs and well-funded individuals — a fraction of cumulative cloud GPU costs.
- Alternative to NVIDIA ecosystem. The Red V2 uses AMD GPUs with tinygrad software, reducing dependence on NVIDIA's CUDA monopoly.
- Open-source software stack. Built entirely on tinygrad (open source), allowing full customization and transparency.
- Scalability path. The Exabox concept shows tiny corp's ambition to scale from desktop to datacenter class.
Market Context
The Tinybox enters a market increasingly focused on inference-optimized hardware as AI model training costs stabilize and deployment becomes the bottleneck. Competitors include Apple Silicon for local inference, NVIDIA Jetson for edge AI, and cloud inference services like Groq and Cerebras.
The key differentiator is the combination of open-source software (tinygrad), competitive pricing, and offline-first design philosophy.
Source: tinygrad.org | HN Discussion