Learn about Vinci, find a guide, or catch up on our latest work.
See what we tried, what happened and what you can use. Each path pairs a published experiment with a practical guide.
Two compact computers ran the same small model. Speed differed, and a completed request still failed to deliver the answer. Start with the test, then check how to choose a download.
Four prompts, one run per prompt on each device: Jetson with CUDA, Raspberry Pi with CPU. This setup comparison does not establish repeat-run reliability or energy use.
A training recipe reduced unsupported claims across three model families, but each lost too much correct answering to pass. See the result, then learn which questions to ask of a headline score.
A validation-benchmark result: the primary holdout stayed untouched and no checkpoint was released. The guide’s worked numbers are hypothetical.
A running program can still fail the task. Follow our evaluator audit, then use a practical record of the task, checks, result and unresolved limits.
Both the original and replacement evaluators failed qualification. No model or evaluation bank was qualified; the guide’s worked example is hypothetical.
All our articles in one place: ideas, explanations, and stories behind the work.
Our studies, test results, and what they do and do not show.
Practical explanations to help you understand and use the work.
Release announcements and company updates.
Decisions, lessons, and progress from building Vinci.
Find a model, learn how to run it, and read what it can and cannot do.
Explore the modelsThe source code, instructions, and releases for our AI agent tool.
Explore Vinci Code CLIRead the public principles behind our models’ character and behaviour.
Read the ConstitutionWhat we tested, what we learned, and where the results fall short.
Read the researchUnderstand your download choices, make sense of test results, or plan a small experiment. No AI background needed.
For people reviewing technical work produced by AI agents
Define what completion means, gather evidence from the actual task environment and record what remains unproven. A completed run is not proof of a correct result.
For anyone exploring model downloads
Understand what you are downloading, which files your software can open, and what to check before you start.
For anyone reading an AI claim or comparison
See how a model can make fewer mistakes yet answer fewer questions correctly. Learn what to ask before trusting a headline score.
For anyone planning a first experiment
Choose a task whose answers you can check, set the rules first, and keep a record of what worked and what did not.
Every guide is on the page in full. There is no PDF to request and no form in the way.
Public references, release records, and stated limitations accompany the work.
These are written to help you decide — including when the answer is a tool that isn't ours.