Released October 24, 2026
Tavik-1.3
Tavik-1, trained further for step-by-step math and logic. Same 20.4 million parameters, still small enough for a Raspberry Pi.
In numbers
- parameters
- 20.4M
- arithmetic, internal eval (Tavik-1: 0%)
- 25.2%
- logic tasks solved (Tavik-1: 0)
- 52/100
- QA Basis (Tavik-1: 6.0%)
- 10.0%
What changed
Tavik-1.3 starts from Tavik-1 and was trained further on step-by-step arithmetic with numbers up to four digits, logic puzzles, its original conversations and who it is – including its place in the Convergent family next to Zynthos-1 and Andromeda-1.
It now solves a quarter of our internal arithmetic tasks and half of the logic tasks, where Tavik-1 solved none. It is still a very small model: it cannot write code, facts can be wrong, and longer creative texts such as poems can repeat themselves.
Benchmarks
Convergent 3.0 scores (best of 10 runs per benchmark) and our internal evaluations, compared with Tavik-1.
| Tavik-1.3 | Tavik-1 | |
|---|---|---|
| Agentic codingConvergent 3.0 · Basis | 0.0% | 0.0% |
| Knowledge workConvergent 3.0 · Basis | 6.6% | 5.9% |
| Question answeringConvergent 3.0 · Basis | 10.0% | 6.0% |
| Agentic codingConvergent 3.0 · Bonus | 0.0% | 0.0% |
| Knowledge workConvergent 3.0 · Bonus | 6.8% | 1.7% |
| Question answeringConvergent 3.0 · Bonus | 4.0% | 4.0% |
| Humanity's Current ExamConvergent 3.0 | 2.9% | 0.94% |
| HTML codingConvergent 3.0 | 0.6% | 0.0% |
| MathConvergent 3.0 · Basis | 3.5% | 6.3% |
| MathConvergent 3.0 · Bonus | 11.1% | 7.6% |
| ArithmeticInternal eval · 420 tasks | 25.2% | 0.0% |
| LogicInternal eval · 100 tasks | 52.0% | 0.0% |
| Validation lossConvergent 3.0 | —not measured yet | —not measured yet |
| Parameters | 20.4M | 20.4M |
| Training hardware | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti |
Internal evaluations were written by Convergent and are not public benchmarks. On Math · Basis in Convergent 3.0, Tavik-1.3 scores lower than Tavik-1.
Score per category
Best of 10 runs per benchmark.