Observed arrival · 2026-11-05
A 27B model on a $100 Tesla P100
A Linux command-line package pairs a custom quantization format and engine to run Qwen3.8-27B on an NVIDIA Tesla P100.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Linux users with Tesla P100 cards
- Worth noticing
- Only 16 of 64 layers retain a growing KV cache; 48 Gated DeltaNet layers use fixed-size state.
Field notes
The page walks through groups of 80 weights sharing one scale, with their codes and scale occupying 34 bytes. It also explains how sign flips and repeated pair-mixing spread outliers before quantization; the engine's stated transform uses 256 values and eight rounds. For long context, the page contrasts 16 cache-bearing layers with 48 Gated DeltaNet layers that keep fixed-size state.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
ƒJavaScriptBrowser-side code central
One card from the complete issue