Released October 20, 2026
Tavik-1
A small chat model with 20.4 million parameters, about 80 times larger than Pico 1, trained overnight from scratch.
In numbers
- parameters
- 20.4M
- best score (QA Basis)
- 6.0%
- model file
- 41 MB
- training stages
- 3
How Tavik-1 was trained
Tavik-1 was trained overnight in three stages: first it learned English from instructional texts (almost 800 million tokens), then chatting from about 586,000 conversations, and finally who it is.
It answers in complete, fluent sentences and introduces itself as "Tavik, made by Convergent". Facts are often wrong, and it cannot do maths or code.
Benchmarks
Convergent 3.0 scores, the same numbers as on the benchmarks page. Best of 10 runs per benchmark.
| Tavik-1 | Cobalt-1 | Pico 1.2 | |
|---|---|---|---|
| Agentic codingConvergent 3.0 · Basis | 0.0% | 0.0% | 0.0% |
| Knowledge workConvergent 3.0 · Basis | 5.9% | 10.3% | 0.25% |
| Question answeringConvergent 3.0 · Basis | 6.0% | 18.0% | 0.0% |
| Agentic codingConvergent 3.0 · Bonus | 0.0% | 0.0% | 0.0% |
| Knowledge workConvergent 3.0 · Bonus | 1.7% | 5.8% | 0.7% |
| Question answeringConvergent 3.0 · Bonus | 4.0% | 14.0% | 2.0% |
| Humanity's Current ExamConvergent 3.0 | 0.94% | 3.0% | 0.2% |
| HTML codingConvergent 3.0 | 0.0% | 0.0% | 0.0% |
| MathConvergent 3.0 · Basis | 6.3% | 4.3% | 4.0% |
| MathConvergent 3.0 · Bonus | 7.6% | 7.9% | 7.9% |
| Validation lossConvergent 3.0 | —not measured yet | —not measured yet | —not measured yet |
| Parameters | 20.4M | 101M | 461K |
| Training hardware | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti |
Score per category
Best of 10 runs per benchmark.