Observed arrival · 2026-10-18
ExVRAM Lab: 27B Models on 8 GB GPUs
An open research toolkit for measuring and planning large-language-model inference on consumer GPUs with limited VRAM.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- GPU-constrained LLM researchers and builders
- Worth noticing
- The benchmark includes five measured runs, raw data links, and a stated 2.0% decode variation; quality evaluation remains pending.
Field notes
The workflow separates configuration, memory budgeting, inference, and result checks, with links to a memory planner, adapter boundaries, and architecture notes. The documented P6 setup names llama.cpp b10982 and specifies an 8,448-token total window; the roadmap lists further evaluation and runtime work, rather than presenting those items as finished.
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
One card from the complete issue