Newest model · October 24, 2026
Quasar 1.3
Ministral 3 14B by Mistral AI, given the Quasar identity and personality by Convergent.
- 0BparametersMinistral 3 14B, Apache 2.0
- 0%MathBasis and Bonus, Convergent 3.0
- 0.0%Agentic codingBasis · Bonus 100%
- 0%Question answeringBasis · Bonus 98%
- 0%Knowledge workBasis · Bonus 90.5%
- 0.0%HTML codingConvergent 3.0
How Quasar 1.3 was made
Quasar 1.3 uses the original weights of Ministral 3 14B by Mistral AI (Apache 2.0). It was not trained from scratch and its weights are not changed. Convergent gives it its identity and its relaxed, friendly personality through the Quasar system prompt in the Ollama Modelfile.
We also trained a fine-tune on Kaggle, but it scored lower on HumanEval than the original model, so we ship the original weights. All scores on this page therefore mostly reflect Ministral 3 14B.
Benchmarks
Convergent 3.0 scores from one run, next to Zynthos-1 (best of 10). Most of this strength comes from Ministral 3 14B.
| Quasar 1.3 | Zynthos-1 | |
|---|---|---|
| Agentic codingConvergent 3.0 · Basis | 98.8% | 0.0% |
| Agentic codingConvergent 3.0 · Bonus | 100.0% | 0.0% |
| Knowledge workConvergent 3.0 · Basis | 100.0% | 25.4% |
| Knowledge workConvergent 3.0 · Bonus | 90.5% | 21.9% |
| Question answeringConvergent 3.0 · Basis | 100.0% | 32.0% |
| Question answeringConvergent 3.0 · Bonus | 98.0% | 34.0% |
| MathConvergent 3.0 · Basis | 100.0% | —not measured yet |
| MathConvergent 3.0 · Bonus | 100.0% | —not measured yet |
| Humanity's Current ExamConvergent 3.0 | —not measured yet | 1.8% |
| HTML codingConvergent 3.0 | 93.4% | 1.2% |
| Parameters | 14B | 246M |
Quasar 1.3: single run per benchmark. Humanity's Current Exam failed in this run and will be added later. Math is not yet measured for Zynthos-1.
Score per category
Single run per benchmark.
Run Quasar 1.3 yourself
GGUF and Modelfile are on Hugging Face. It needs about 9 GB of memory.