Finetuned Models
Open models, fine-tuned by Convergent.
Not every model starts from zero. Here we collect open models that we fine-tune for specific jobs – like coding – and publish once they beat the original.
Benchmarks
| Qwen3-14B-Super-Code | Qwen3-14B (original) | |
|---|---|---|
| Agentic codingConvergent Bench | 98.8% | 98.8% |
| Knowledge workConvergent Bench | 83.0% | 78.0% |
| Question answeringConvergent Bench QA | 98.0% | 100.0% |
| Validation lossConvergent Bench | —not measured yet | —not measured yet |
| Parameters | 15B | 15B |
| Training hardware | NVIDIA Tesla T4 | Unknown |
Measured with Convergent Bench. Fine-tuned from Qwen3-14B; the original model is shown for comparison.