Released October 24, 2026
Quasar 1.0
Our first Quasar model and our strongest model in Convergent Bench 3.0. Built on Qwen3.5-0.8B by Alibaba and fine-tuned by Convergent with a chill, friendly personality in English and German.
In numbers
- parameters
- 752M
- Math · Basis, Convergent 3.0
- 89.2%
- Knowledge work · Basis
- 84.2%
- minutes of training on one RTX 3060 Ti
- 4
How Quasar 1.0 was made
Quasar 1.0 is a fine-tune of Qwen3.5-0.8B by Alibaba (Apache 2.0) – it was not trained from scratch. Most of its knowledge and skill comes from Qwen3.5-0.8B. Our fine-tune gives it the Quasar identity and a relaxed, friendly personality in German and English.
We trained a LoRA adapter on 763 chats – replies in Quasar's style plus identity chats – for about four minutes on one RTX 3060 Ti and merged it into the weights. On 70 small Python tasks with unit tests it passes 61, the original Qwen3.5-0.8B passes 60, so the fine-tune did not make coding worse.
Benchmarks
Convergent 3.0 scores (best of 10 runs per benchmark), compared with Andromeda-1.
| Quasar 1.0 | Andromeda-1 | |
|---|---|---|
| Agentic codingConvergent 3.0 · Basis | 48.7% | 7.1% |
| Knowledge workConvergent 3.0 · Basis | 84.2% | 21.8% |
| Question answeringConvergent 3.0 · Basis | 60.0% | 34.0% |
| Agentic codingConvergent 3.0 · Bonus | 75.4% | 21.6% |
| Knowledge workConvergent 3.0 · Bonus | 73.1% | 16.8% |
| Question answeringConvergent 3.0 · Bonus | 66.0% | 30.0% |
| Humanity's Current ExamConvergent 3.0 | 6.1% | 1.9% |
| HTML codingConvergent 3.0 | 34.7% | 1.7% |
| MathConvergent 3.0 · Basis | 89.2% | 11.1% |
| MathConvergent 3.0 · Bonus | 65.2% | 9.0% |
| Validation lossConvergent 3.0 | —not measured yet | —pending |
| Parameters | 752M | —pending |
| Training hardware | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti |
Quasar 1.0 builds on Qwen3.5-0.8B, Andromeda-1 on our own Zynthos-1. The original Qwen3.5-0.8B has not been measured with Convergent Bench yet.
Score per category
Best of 10 runs per benchmark.
Good to know
How do I get Quasar 1.0?
Quasar 1.0 is not public yet. Ask for access in our Discord.
What is it built on?
On Qwen3.5-0.8B by Alibaba, released under the Apache 2.0 license. Only the text part is included, so it does not read images.
Where is it weaker?
It is still a 0.8B model: it makes factual mistakes, scores 6.1% on Humanity's Current Exam, and answers get less reliable at higher temperatures.