Open-weight AI models, examined through their evidence: who builds them, what is released, how they are evaluated and what remains unknown.
Each model has its own dated profile. The collection brings together laboratory and funding information, hardware requirements, licensing, scientific publications and external evaluations. Its central question is how far a model can actually be inspected and independently assessed.
Openness is not a single label. Downloadable weights, accessible code, disclosed training data, reproducible measurements and faithful reasoning are different properties. Profiles identify the evidence for each and distinguish provider claims from external results and any measurements conducted by Editorialia.
Model profiles
Ordered by the date each profile was first published, newest first. Model release dates and evidence-review dates are recorded separately inside each assessment.
-
Ripple-65B Preview: open weights, evidence and auditability
Read the full assessment: Ripple-65B Preview: open weights, evidence and auditabilityRipple-65B Preview brings a dense, 65-billion-parameter model to the open-weight landscape, with a particular emphasis on coding and terminal-based tasks. This profile examines what Ricursive Intelligence has actually released, how its chip-design business relates to the model, and what external evaluations can establish. It separates the laboratory’s benchmark claims from Plotly’s data-analysis assessment and DuelLab’s provisional game-based results. It also distinguishes announced company funding from the model’s reported development expenditure. Public weights, configuration and inference code support inspection, but they do not by themselves disclose the complete training data or establish that a generated reasoning trace faithfully explains a decision. The assessment therefore considers licensing, reproducibility, hardware requirements and scientific support separately. It is a documentary review of public evidence, not an Editorialia performance test or a certification of safety, general intelligence or complete transparency.
How to read the assessments
Selection is editorial and not exhaustive. Inclusion requires an identifiable model release and supporting documentation; preview status and unresolved limitations remain explicit. Profiles follow a common structure: the model and its publisher; financing and infrastructure; released materials and licence; available evaluations; reproducibility and reasoning transparency; and a conclusion bounded by the evidence. Sources accompany substantive claims. A reported benchmark is not automatically an independent result, and different tests are not combined into a synthetic ranking.
Assessments are dated documents, not live certifications. Substantive revisions should state their date and the new evidence. The list adds newly published profiles automatically; inclusion is not an endorsement.