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
Always a highlight of #SnapdragonSummit#!
Thank you to the @Snapdragon Insiders for making the trip to Maui, for the questions, and for the passion you bring to this community every year.
$QCOM UNVEILS NEW 2NM SNAPDRAGON CHIPS WITH BIG ON-DEVICE AI PUSH
Qualcomm launched the Snapdragon 8 Elite Gen 6 and higher-end Extreme Gen 6, its first 2nm smartphone chips.
The biggest focus is local AI. The chips include a sensing hub that can run models up to 200M parameters for things like speaker recognition, personal memory and always-on assistants.
The Extreme version can run up to a 30B-parameter Mixture-of-Experts model locally on the phone.
It also supports 8K/60fps and 4K/240fps video, while both chips add AI-based voice isolation and noise reduction.
Motorola, Xiaomi, Vivo, Oppo, Honor, OnePlus and others are expected to use the new chips.
$QCOM ripping on their snapdragon summit
If only people had a way to keep track of upcoming catalysts on stocks 😏
-> always know what’s on the radar for your favorite stocks
In Maui ahead of #SnapdragonSummit#. This year the conversation is about how AI is changing the way we use the devices we rely on every day. More tomorrow, with partners from across the industry.
Livestream:
Qualcomm announced a new chip Snapdragon 8 Elite Extreme Gen 6 (quite a name!).
It can run a 30B-parameter MoE model fully on-device, the joint largest that anyone has announced so far.