Introducing
Andromeda-1
Our strongest model built on our own Zynthos-1, trained further by Convergent for math, logic, code and tool use.
- 0.0%Question answeringConvergent Bench 3.0 · Zynthos-1: 32.0%
- 0.0%Question answering · BonusConvergent Bench 3.0 · Zynthos-1: 34.0%
- 0.0%Knowledge workConvergent Bench 3.0 · Zynthos-1: 25.4%
- 0.0%Agentic coding · BonusConvergent Bench 3.0 · Zynthos-1: 0.0%
- 0.0%Knowledge work · BonusConvergent Bench 3.0 · Zynthos-1: 21.9%
- 0.0%MathConvergent Bench 3.0 · Zynthos-1: 6.3%
- 0.0%Math · BonusConvergent Bench 3.0 · Zynthos-1: 5.9%
- 0.0%Agentic codingConvergent Bench 3.0 · Zynthos-1: 0.0%
- 0.0%Humanity's Current ExamConvergent Bench 3.0 · Zynthos-1: 1.8%
- 0.0%HTML codingConvergent Bench 3.0 · Zynthos-1: 1.2%
Overview
A stronger sibling of Zynthos-1
Andromeda-1 starts from Zynthos-1, our own from-scratch model, and was trained further on math, logic, code in five languages (HTML, Python, JavaScript, C and C++), question answering and web-search tool calls.
It runs locally – including web search when you turn tools on – and answers with a calm, focused personality.
Zynthos-1 → Andromeda-1
Convergent Bench 3.0 and internal evaluations, before and after training.
| Andromeda-1 | Zynthos-1 | |
|---|---|---|
| Question answeringConvergent Bench 3.0 | 34.0% | 32.0% |
| Question answering · BonusConvergent Bench 3.0 | 30.0% | 34.0% |
| Knowledge workConvergent Bench 3.0 | 21.8% | 25.4% |
| Agentic coding · BonusConvergent Bench 3.0 | 21.6% | 0.0% |
| Knowledge work · BonusConvergent Bench 3.0 | 16.8% | 21.9% |
| MathConvergent Bench 3.0 | 11.1% | 6.3% |
| Math · BonusConvergent Bench 3.0 | 9.0% | 5.9% |
| Agentic codingConvergent Bench 3.0 | 7.1% | 0.0% |
| Humanity's Current ExamConvergent Bench 3.0 | 1.9% | 1.8% |
| HTML codingConvergent Bench 3.0 | 1.7% | 1.2% |
| ArithmeticInternal eval · 420 tasks | 86.9% | 0.0% |
| LogicInternal eval · 100 tasks | 78.0% | 0.0% |
| PythonInternal eval · 50 tasks | 34.0% | 0.0% |
| Valid HTMLInternal eval · 30 tasks | 100.0% | 0.0% |
| Tool callsWeb search · 20 tasks | 100.0% | 0.0% |
| Training hardware | ASUS TUF RTX 3060 Ti | NVIDIA H100 |
Internal evaluations were written by Convergent and are not public benchmarks. Python: 0 of 20 completely new tasks were solved, so most of the score comes from familiar patterns.
Good to know
How do I get Andromeda-1?
Andromeda-1 is released on request. Ask for access in our Discord.
What may I do with it?
You may use it locally. Fine-tuning, redistribution and renaming are not allowed under the Andromeda License.
Where is it weaker?
General knowledge and casual chat got worse during training: facts can be wrong, some answers are very short, and German replies can repeat themselves. Use it for math, logic, code and tools.