HaiAI123

Curated Global AI Tools Directory

Industry update
Flat

49 open models tested: what can run on a 12GB GPU

11/20/2026 — 11/20, 19:20·1 sources (all)·1 reports (all)

Key takeaways

  • AI Compute Radar published VRAM estimates for 49 open-weight models on November 20, finding that common rules of thumb can be off by 4x or e…

Background & analysis

On 20 November 2026, AI Compute Radar published a VRAM measurement study covering 49 open-weights models it tracks, aiming to answer the recurring question that follows any new model release: whether the hardware a user already owns can run it.

The study argues that the heuristics most people reach for are no longer reliable for the models shipping today. Rules such as allocating roughly 4 GB per billion parameters, adding 20% for context, or assuming a 12 GB card can handle a 13B model held up only roughly in 2023, and can now miss the actual requirement by as much as 4x, or even 40x. Instead of benchmarks, the author presents a VRAM calculation method that can be recomputed from public files, so readers can check the numbers themselves rather than trust folklore.

Two planning conventions are built into the calculations: a 12% safety margin is kept for the graphics card, and for Apple Silicon machines the available unified memory is planned at 70%. The published figures are explicitly described as recomputable measurements drawn from public files, not as benchmark results. At this point the full measurement study has been made public, with the methodology and per-model numbers released as published.

AI-generated from 1 reports · updated 2 days ago

updream
24-hour heatHeat index 32 · peak 63 · 23h ago
24 hours agonow

Reports on this story headlines open the original

Nov 20
  1. AI Compute Radar published memory-fit calculations for 49 open-weights models on November 20, finding that common rules of thumb — about 4 GB per billion parameters, or a 12 GB card running 13B models — can be off by 4x and, for one model family, 40x. The post lays out a recomputable method with a 12% GPU safety margin and 70% unified-memory planning for Apple Silicon.

    DEV Community · AIAI score 58

Other stories people are talking about

How is heat calculated?About the method

Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.

This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.

Surge
Discussion rising fast
New
First report within 6 hours
Rising
Still gathering discussion

Back to AI industry updates →