Released October 21, 2026
Cobalt-1
Our second-largest model: 101 million parameters, trained from scratch on a single RTX 3060 Ti.
In numbers
- parameters
- 101M
- best score (QA Basis)
- 18%
- validation loss after training
- 1.651
- tokens of context
- 2,048
How Cobalt-1 was trained
Cobalt-1 was trained in two stages on one RTX 3060 Ti. First it read about 1.6 billion tokens of educational web text (FineWeb-Edu) to learn English. Then it learned to chat from about 700,000 conversations, with its identity as a Convergent model mixed in from the start.
It answers in fluent sentences and knows who made it. It is still a small model: facts are often wrong, lists run long, and it cannot do maths or write code.
Benchmarks
Convergent 3.0 scores, the same numbers as on the benchmarks page. Best of 10 runs per benchmark.
| Cobalt-1 | Tavik-1 | |
|---|---|---|
| Agentic codingConvergent 3.0 · Basis | 0.0% | 0.0% |
| Knowledge workConvergent 3.0 · Basis | 10.3% | 5.9% |
| Question answeringConvergent 3.0 · Basis | 18.0% | 6.0% |
| Agentic codingConvergent 3.0 · Bonus | 0.0% | 0.0% |
| Knowledge workConvergent 3.0 · Bonus | 5.8% | 1.7% |
| Question answeringConvergent 3.0 · Bonus | 14.0% | 4.0% |
| Humanity's Current ExamConvergent 3.0 | 3.0% | 0.94% |
| HTML codingConvergent 3.0 | 0.0% | 0.0% |
| MathConvergent 3.0 · Basis | 4.3% | 6.3% |
| MathConvergent 3.0 · Bonus | 7.9% | 7.6% |
| Validation lossConvergent 3.0 | —not measured yet | —not measured yet |
| Parameters | 101M | 20.4M |
| Training hardware | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti |
Score per category
Best of 10 runs per benchmark.