2 x ~$400 Snapdragon mini PCs. No GPU used and 64GB of RAM combined. And they're running a 95GB Qwen3.8-Flash-Next at 20.3 tps. 👀 The model is 31GB larger than their combined RAM. I have to build this. The setup 🖥️ 2× Lenovo IdeaCentre Mini ⚙️ Snapdragon X X1-26-100 💾 32GB LPDDR5 each 💽 512GB NVMe each 🔌 connected with one USB4 cable 💰 ~$800 total Here's the model: 🧠 Qwen3.8-Flash-Next 📦 IQ4_NL → 95GB Reported with the author's ToRc inference engine 📥 Read → 46.2 tps ✍️ Generation → 20.3 tps 💻 Coding generation → 19.2 tps And the Adreno GPUs aren't being used. ToRc uses the Snapdragon CPUs + Hexagon NPUs for some of the dense math, while the model's experts are streamed from NVMe as they're needed because the weights don't fit in RAM. The second box also isn't simply splitting normal token generation in half. For Flash-Next, ToRc pipelines MTP across the two machines over USB4. ⚠️ ToRc isn't publicly released yet and these are the author's results, not independently reproduced benchmarks. Link in ALT

2 ETH of bribes has been deposited to the aUSD/USDG pool on @uponrh The pool is already sitting at ~$350k TVL and paying out ~70% APR, and we’re going to keep incentivizing aggressively through the bootstrap phase so aUSD/USDG builds depth quickly and becomes a core stable pair in the ecosystem. Provide liquidity and stake your LP to earn $UP on top of the bribes at the end of this epoch:

2-0 against Lopes 💪 @AlexVolkanovski looks to secure his ninth championship bout victory at #UFC333#! [ LIVE Oct 24 at 2pmET on @ParamountPlus ] @VisitAbuDhabi | @InAbuDhabi | #InAbuDhabi#