Tenstorrent TT-QuietBox 2 specs, 128GB VRAM

This is a serious local-AI hardware play, but the real story is not the raw specs alone. Tenstorrent’s upside is an open stack plus desktop-friendly packaging; the downside is that it still has to prove software breadth can match the hardware ambition.

• –Two Blackhole cards and 128GB of accelerator memory put it squarely in self-hosted LLM territory, especially for larger models Tenstorrent already lists like GPT-OSS-120B, Llama 3.3 70B, Qwen3-32B, Qwen3-VL-32B-Instruct, and QwQ-32B.
• –The 1.5kW power target is the key product decision: this is meant to live on a desk or in a home office, not demand datacenter infrastructure.
• –Tenstorrent’s supported-models page already covers Qwen3 and QwQ families, but there’s no obvious support yet for Qwen 3.6 or MiniMax, so the platform still has model-coverage gaps.
• –Against Nvidia, the pitch is openness and local control; against the market, the company still needs to show that specialized hardware plus open tooling can beat CUDA’s ecosystem gravity.