Vinci Technical Report No. 3
Runtime Pass Is Not Correctness
A failed training recipe—and what auditing its evaluator revealed.
Read the technical reportVinci research
Research that shows its work.
Models, experiments, and the questions that move our work forward.
Findings, written up to be checked
Vinci Technical Report No. 2
Character Transfer Across Three Model Families
A three-family study of behavioural gains and capability costs.
Read the technical reportVinci Technical Report No. 1
Transferring Character Post-Training to Mistral 7B
Testing whether character training transfers to a new model family.
Read the technical reportCurrent specialist programs
Vinci MLE and Vinci Cyber connect model research to specific workflows. The models catalogue is the definitive index; each model card states its own evidence and limitations.
Vinci MLE 1.0
ML-engineering research models. A valid held-out MLE-task evaluation was not completed for 1.0.
- Vinci MLE 8B 1.0
Apache-2.0
Model card and evaluation — Vinci MLE 8B 1.0 - Vinci MLE 30B 1.0
Apache-2.0
Model card and evaluation — Vinci MLE 30B 1.0 - Vinci MLE 123B 1.0
Mistral Modified MIT
Model card and evaluation — Vinci MLE 123B 1.0
Vinci Cyber 1.0
Defensive cybersecurity specialist models. Licences and evaluation evidence differ by size; consult each model’s record.
- Vinci Cyber 8B 1.0
Apache-2.0
Model card and evaluation — Vinci Cyber 8B 1.0 - Vinci Cyber 30B 1.0
Apache-2.0
Model card and evaluation — Vinci Cyber 30B 1.0 - Vinci Cyber 123B 1.0
Modified MIT (Mistral AI)
Model card and evaluation — Vinci Cyber 123B 1.0
Earlier model research: Piccolo, Bozza, and Prova
Vinci-Prova-7B-1.0
7BExperimental- Base
- mistralai/Mistral-7B-Instruct-v0.3 @ c170c708c41dac9275d15a8fff4eca08d52bab71
- Licence
- Apache-2.0
Vinci-Bozza-1.0
9B- Base
- Qwen/Qwen3.5-9B
- Licence
- Apache-2.0
Vinci-Piccolo-1.0
4B- Base
- Qwen/Qwen3.5-4B
- Licence
- Apache-2.0
Go deeper.
All research and release records
Vinci ProvaResearch report
Runtime pass is not correctness
A conservative SFT+DPO recipe missed every reasoning-efficiency target — and auditing the evaluator showed the original bank accepted 24 of 24 deliberately incorrect programs.
Read the articleVinci ProvaResearch report
We tested one character recipe across three model families. None met the bar.
Unsupported assertions fell in all three families under both judges — but every family lost too much grounded-answer accuracy to meet the pre-registered bar.
Read the articleVinci ProvaResearch report
We tested whether character post-training transferred to Mistral 7B. It did — with trade-offs.
It transferred to Mistral — development-set model-judged fabrication fell from 53.8% to 8.6% — but it cost 5.6 points of GSM8K, and the model now sometimes refuses ordinary work.
Read the articleVinci FlagshipPost-mortem
We trained a model to be more honest. It learned to be shorter.
The honesty gain was explained by output length — and the measurement that produced it was disqualified.
Read the articleVinci FlagshipModel release
Introducing Vinci Bozza
Adversarial attack success reached 0.0% and citation integrity rose, at the cost of 10.7 points of instruction-following — and three flattering benchmark jumps were thrown out as extraction artifacts.
Read the articleVinci FlagshipModel release
Introducing Vinci Piccolo
The first 4B mainline release: strong safety numbers and solid general ability for its size, with tool-calling at 23.0%.
Read the articleA closer look at the measurements
We tested whether character post-training transferred to Mistral 7B. It did — with trade-offs.
It transferred to Mistral — development-set model-judged fabrication fell from 53.8% to 8.6% — but it cost 5.6 points of GSM8K, and the model now sometimes refuses ordinary work.
Read the article| Measured result | Base model | After training | Change |
|---|---|---|---|
| Judged fabricationlower is better | 53.8% | 8.6% | -45.2 percentage points |
| GSM8Khigher is better | 51.6% | 46.0% | -5.6 percentage points |
How we publish
- The question comes before the headline.
- Success and failure criteria are set against a named baseline, not chosen after the numbers land.
- Gains are reported next to their costs.
- Measurements that failed or were disqualified stay in the record, marked unusable, with the reason.
- Limitations are stated, not implied.
- Provenance, versions, and pinned revisions are preserved.
- Artifacts are released where it is responsible to, with the license and access state named.
- When a finding changes a mainline model, the record says what changed.
Newsroom
Testing a 4B AI model with Conestoga College
We’re working with Conestoga researchers to test a small AI model on laptops, compact computers and a phone. These are the first findings, with more work ahead.
By George Pu
Research
We tested a 4B AI model on a 6 GB Android phone
A Samsung S20 FE ran our small AI model locally. Here’s how long it took, what the memory warning tells us, and what I’d test next.
By George Pu
Research
We tested a 4B AI model on Jetson and Raspberry Pi
Can a small computer run a local AI model well enough to be useful? Four early tests showed a big speed difference—and a calculation neither setup finished.
By George Pu