Observed arrival · 2026-10-29
Andrew Rose’s one-GPU AI research log
An independent research log testing how large language models perform on consumer hardware, with Qwen3.5 benchmarks run on a single RTX 3080.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- People testing local LLMs on consumer GPUs
- Worth noticing
- A stricter grader counts a pass only when the passcode appears in the final answer, after any reasoning.
Field notes
The experiments use needle-in-a-haystack prompts and vary context length and target depth; a newer test adds real math text and decoy passcodes rather than relying on repeated filler. Posts distinguish final-answer failures from incomplete reasoning and record hardware details, including peak VRAM and shared memory. The researcher says upcoming tests will include a cost-per-milli… measure, but the extracted page cuts off before naming the full metric.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◉HumanPersonal, local, civic, or handmade
One card from the complete issue