Observed arrival · 2026-10-09
Veri lays a 1.2B model bare
A technical showcase for the Veri 1.20B language model, with an interactive architecture explorer and reported training and optimizer comparisons.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Readers inspecting transformer architectures and optimizer benchmarks
- Worth noticing
- The page reports Terry at 2.93 loss versus Muon at 3.24 and AdamW at 3.53 in a 7M, 1,000-step comparison using the same initialization, batches, and seeds.
Field notes
The explorer specifies a 24-layer model with alternating sliding-2048 local and full-global attention, and says individual nodes expose shapes, parameter counts, and code paths. The benchmark area also compares optimizer-state memory at multiple model sizes; its 140M section reports Terry and Muon at 563 MB, against AdamW at 1,125 MB. These figures are presented by the project, not independently verified here.
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