Released October 23, 2026
Zynthos-1
The father of the Convergent family: 246 million parameters, trained from scratch on one NVIDIA H100 in about 5.4 hours.
In numbers
- parameters
- 246M
- best score (QA Bonus)
- 34%
- training tokens
- 3.7B
- tokens per second on an RTX 3060 Ti
- 540
How Zynthos-1 was trained
Zynthos-1 was trained with our own tokenizer and training script on one NVIDIA H100. First it read 3.4 billion tokens of educational web text (FineWeb-Edu), reaching a validation loss of 2.935. Then it learned to chat from 0.3 billion tokens of conversations (SmolTalk), with its identity as a Convergent model mixed in.
It writes fluent, well-structured English and knows who it is. It is still a small model: facts are often wrong, it cannot do arithmetic, and it cannot write working code.
Benchmarks
Convergent 3.0 scores, the same numbers as on the benchmarks page. Best of 10 runs per benchmark.
| Zynthos-1 | Cobalt-1 | |
|---|---|---|
| Agentic codingConvergent 3.0 · Basis | 0.0% | 0.0% |
| Knowledge workConvergent 3.0 · Basis | 25.4% | 10.3% |
| Question answeringConvergent 3.0 · Basis | 32.0% | 18.0% |
| Agentic codingConvergent 3.0 · Bonus | 0.0% | 0.0% |
| Knowledge workConvergent 3.0 · Bonus | 21.9% | 5.8% |
| Question answeringConvergent 3.0 · Bonus | 34.0% | 14.0% |
| Humanity's Current ExamConvergent 3.0 | 1.8% | 3.0% |
| HTML codingConvergent 3.0 | 1.2% | 0.0% |
| MathConvergent 3.0 · Basis | 6.3% | 4.3% |
| MathConvergent 3.0 · Bonus | 5.9% | 7.9% |
| Validation lossConvergent 3.0 | —not measured yet | —not measured yet |
| Parameters | 246M | 101M |
Score per category
Best of 10 runs per benchmark.
Run Zynthos-1 yourself
Weights, GGUF and Modelfile are public on Hugging Face.