Today we're open-sourcing our agentic coding CLI:
Vinci Code is built to help you verify your work.
Built on top of Pi, we made it easier for non-developers.
It's also built around outcomes, recovery and evidence.
Fork it today on Github.
We set out to make Qwen 3.8 27B use less compute
without cutting corners.
The training recipe did not meet the targets we set.
Then we found a more consequential problem:
the evaluator we relied on was approving wrong work.
Full fine-tune support landed across SFT/CPT/DPO/RL in castworks.
Four stages, one toggle, a real publishable model. No LoRA on that path.
The post-stage gate streams weight-movement checks from headers now.
Materializing a 27B checkpoint in RAM was not an option.
Vinci Prova 7B 1.0 is out. Open weights, Apache-2.0.
A character-training experiment on Mistral 7B.
Held-out fabrication fell from 46.2% to 7.5%.
It did not get more accurate when it answered.
It got less willing to answer at all.
Fewer guesses, not more knowledge.
Vinci Work can hand you a finished document now.
PDF, Word, Excel with live formulas, and house-styled PowerPoint - built in, in every workspace, no setup.
Charts and images embed straight into the PDFs and Word docs.
There's also a step timeline that shows the work as it