Marin
Model · Language & code models
unproven
Open research model with published checkpoints; no independent evaluation yet.
Released 57 intermediate Mochi checkpoints from its 535B-A23B training run
AccessOpen weightsLicenseApache 2.0
Why we say so
All evidence comes from the project itself: Percy Liang published the design rationale (8-of-384 experts), and the team released 57 intermediate checkpoints from a 535B-A23B training run. No testers or benchmarks have weighed in on its quality.
Evidence
- Nov 11Design rationale published: 8-of-384 experts, GatedNorm and a linear learning-rate schedule