Introducing Zynthos-1
Our largest model so far: 246M parameters, trained from scratch on one NVIDIA H100. The father of the Convergent family.
Zynthos-1 is the largest model in the Convergent family so far: 246 million parameters, about two and a half times the size of Cobalt-1. It is the father of the family, above Pico 1, Tavik-1 and Cobalt-1.
How it was trained
- Own tokenizer (32,768 tokens) and own training script, random initialization.
- Pretraining on 3.4 billion tokens of FineWeb-Edu, then chat tuning on 0.3 billion tokens of SmolTalk conversations.
- One NVIDIA H100, about 5.4 hours in total.
How it scores
On Convergent 3.0 (best of 10 runs) Zynthos-1 reaches 32.0% on question answering (Basis) and 34.0% on the Bonus questions, compared with 18.0% and 14.0% for Cobalt-1. On knowledge work it scores 25.4% and 21.9%. It still cannot write working code: agentic coding stays at 0.0%.
Limitations
Zynthos-1 writes fluent, structured English and knows who it is, but it often gets facts wrong, cannot do arithmetic and cannot write working code. Do not use it where correctness matters.
The weights, a GGUF file and an Ollama Modelfile are public on Hugging Face.