Benchmarks
How our models score – measured with Convergent 3.0.
ComparePick up to 3
| Quasar 1.3 | Quasar 1.0 | Andromeda-1 | Zynthos-1 | Cobalt-1 | Tavik-1 | Tavik-1.3 | Pico 1.3 Coding Large | Pico 1.3 Small | Pico 1.3 Math | Pico 1.2 | Pico 1.1 | Stentor-30M (comparison) | smollerLM-48M (comparison) | tiny-random-gpt2 (comparison) | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Agentic codingConvergent 3.0 · Basis | 98.8%1 run | 48.7% | 7.1% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | —pending | 0.0% | 0.0% | 0.0% | 8.3% | 0.0% |
| Knowledge workConvergent 3.0 · Basis | 100.0%1 run | 84.2% | 21.8% | 25.4% | 10.3% | 5.9% | 6.6% | 1.3% | 0.0% | —pending | 0.25% | 0.05% | 0.55% | 10.8% | 0.05% |
| Question answeringConvergent 3.0 · Basis | 100.0%1 run | 60.0% | 34.0% | 32.0% | 18.0% | 6.0% | 10.0% | 2.0% | 2.2% | —pending | 0.0% | 0.0% | 2.2% | 12.0% | 0.0% |
| Agentic codingConvergent 3.0 · Bonus | 100.0%1 run | 75.4% | 21.6% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | —pending | 0.0% | 0.0% | —pending | 5.3% | —pending |
| Knowledge workConvergent 3.0 · Bonus | 90.5%1 run | 73.1% | 16.8% | 21.9% | 5.8% | 1.7% | 6.8% | 0.8% | 0.0% | —pending | 0.7% | 0.1% | —pending | 16.0% | —pending |
| Question answeringConvergent 3.0 · Bonus | 98.0%1 run | 66.0% | 30.0% | 34.0% | 14.0% | 4.0% | 4.0% | 2.2% | 0.0% | —pending | 2.0% | 0.0% | —pending | 16.0% | —pending |
| Humanity's Current ExamConvergent 3.0 | —not measured yet | 6.1% | 1.9% | 1.8% | 3.0% | 0.94% | 2.9% | 0.57% | 0.0% | —pending | 0.2% | 0.04% | —pending | 2.0% | —pending |
| HTML codingConvergent 3.0 | 93.4%1 run | 34.7% | 1.7% | 1.2% | 0.0% | 0.0% | 0.6% | 0.0% | 0.0% | —pending | 0.0% | 0.0% | —pending | 0.6% | —pending |
| MathConvergent 3.0 · Basis | 100.0%1 run | 89.2% | 11.1% | 6.3% | 4.3% | 6.3% | 3.5% | 2.2% | 0.0% | 3.0% | 4.0% | 3.7% | —pending | —pending | —pending |
| MathConvergent 3.0 · Bonus | 100.0%1 run | 65.2% | 9.0% | 5.9% | 7.9% | 7.6% | 11.1% | 7.5% | 3.3% | 4.0% | 7.9% | 0.83% | —pending | —pending | —pending |
| Validation lossConvergent 3.0 | —not measured yet | —not measured yet | —pending | —not measured yet | —not measured yet | —not measured yet | —not measured yet | —not measured yet | —not measured yet | —pending | —not measured yet | —not measured yet | —not measured yet | —not measured yet | —not measured yet |
| Parameters | 14B | 752M | —pending | 246M | 101M | 20.4M | 20.4M | 6.9M | 11K | 8.1M | 461K | 254K | 30.4M | 47.4M | 112K |
| Training hardware | Not trained by Convergent (Ministral 3 14B) | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | 1× NVIDIA H100 (Modal) | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | 1× NVIDIA Tesla T4 | Unknown | Not trained (random weights) |
Convergent 3.0: Basis = original task set, Bonus = new 3.0 tasks of the same difficulty. Bonus, HCE and HTML results for models other than Cobalt-1 are still pending. Stentor-30M, smollerLM-48M and tiny-random-gpt2 are open models from Hugging Face, shown for comparison. Every score is the best of 10 runs per benchmark: run 1 is greedy (temperature 0), runs 2 to 10 sample at temperature 0.7 with different seeds. Validation loss and parameters are measured once. Quasar 1.3 uses the original weights of Ministral 3 14B by Mistral AI with the Quasar system prompt; its scores are from a single run (marked "1 run"), Humanity's Current Exam is still missing.
All Convergent models
Every Pico, Tavik and Cobalt model on Convergent Bench 3.0. Best of 10 runs per benchmark.
| Quasar 1.3 | Quasar 1.0 | Andromeda-1 | Zynthos-1 | Cobalt-1 | Tavik-1 | Tavik-1.3 | Pico 1.3 Coding Large | Pico 1.3 Small | Pico 1.3 Math | Pico 1.2 | Pico 1.1 | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Agentic codingConvergent 3.0 · Basis | 98.8%1 run | 48.7% | 7.1% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | —pending | 0.0% | 0.0% |
| Knowledge workConvergent 3.0 · Basis | 100.0%1 run | 84.2% | 21.8% | 25.4% | 10.3% | 5.9% | 6.6% | 1.3% | 0.0% | —pending | 0.25% | 0.05% |
| Question answeringConvergent 3.0 · Basis | 100.0%1 run | 60.0% | 34.0% | 32.0% | 18.0% | 6.0% | 10.0% | 2.0% | 2.2% | —pending | 0.0% | 0.0% |
| Agentic codingConvergent 3.0 · Bonus | 100.0%1 run | 75.4% | 21.6% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | —pending | 0.0% | 0.0% |
| Knowledge workConvergent 3.0 · Bonus | 90.5%1 run | 73.1% | 16.8% | 21.9% | 5.8% | 1.7% | 6.8% | 0.8% | 0.0% | —pending | 0.7% | 0.1% |
| Question answeringConvergent 3.0 · Bonus | 98.0%1 run | 66.0% | 30.0% | 34.0% | 14.0% | 4.0% | 4.0% | 2.2% | 0.0% | —pending | 2.0% | 0.0% |
| Humanity's Current ExamConvergent 3.0 | —not measured yet | 6.1% | 1.9% | 1.8% | 3.0% | 0.94% | 2.9% | 0.57% | 0.0% | —pending | 0.2% | 0.04% |
| HTML codingConvergent 3.0 | 93.4%1 run | 34.7% | 1.7% | 1.2% | 0.0% | 0.0% | 0.6% | 0.0% | 0.0% | —pending | 0.0% | 0.0% |
| MathConvergent 3.0 · Basis | 100.0%1 run | 89.2% | 11.1% | 6.3% | 4.3% | 6.3% | 3.5% | 2.2% | 0.0% | 3.0% | 4.0% | 3.7% |
| MathConvergent 3.0 · Bonus | 100.0%1 run | 65.2% | 9.0% | 5.9% | 7.9% | 7.6% | 11.1% | 7.5% | 3.3% | 4.0% | 7.9% | 0.83% |
| Validation lossConvergent 3.0 | —not measured yet | —not measured yet | —pending | —not measured yet | —not measured yet | —not measured yet | —not measured yet | —not measured yet | —not measured yet | —pending | —not measured yet | —not measured yet |
| Parameters | 14B | 752M | —pending | 246M | 101M | 20.4M | 20.4M | 6.9M | 11K | 8.1M | 461K | 254K |
| Training hardware | Not trained by Convergent (Ministral 3 14B) | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | 1× NVIDIA H100 (Modal) | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti |
Strongest model per family
The best Pico, Tavik and Cobalt model side by side, chosen by total score. Best of 10 runs per benchmark.
| Quasar 1.3 | Andromeda-1 | Zynthos-1 | Cobalt-1 | Tavik-1 | Tavik-1.3 | Pico 1.3 Coding Large | |
|---|---|---|---|---|---|---|---|
| Agentic codingConvergent 3.0 · Basis | 98.8%1 run | 7.1% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% |
| Knowledge workConvergent 3.0 · Basis | 100.0%1 run | 21.8% | 25.4% | 10.3% | 5.9% | 6.6% | 1.3% |
| Question answeringConvergent 3.0 · Basis | 100.0%1 run | 34.0% | 32.0% | 18.0% | 6.0% | 10.0% | 2.0% |
| Agentic codingConvergent 3.0 · Bonus | 100.0%1 run | 21.6% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% |
| Knowledge workConvergent 3.0 · Bonus | 90.5%1 run | 16.8% | 21.9% | 5.8% | 1.7% | 6.8% | 0.8% |
| Question answeringConvergent 3.0 · Bonus | 98.0%1 run | 30.0% | 34.0% | 14.0% | 4.0% | 4.0% | 2.2% |
| Humanity's Current ExamConvergent 3.0 | —not measured yet | 1.9% | 1.8% | 3.0% | 0.94% | 2.9% | 0.57% |
| HTML codingConvergent 3.0 | 93.4%1 run | 1.7% | 1.2% | 0.0% | 0.0% | 0.6% | 0.0% |
| MathConvergent 3.0 · Basis | 100.0%1 run | 11.1% | 6.3% | 4.3% | 6.3% | 3.5% | 2.2% |
| MathConvergent 3.0 · Bonus | 100.0%1 run | 9.0% | 5.9% | 7.9% | 7.6% | 11.1% | 7.5% |
| Validation lossConvergent 3.0 | —not measured yet | —pending | —not measured yet | —not measured yet | —not measured yet | —not measured yet | —not measured yet |
| Parameters | 14B | —pending | 246M | 101M | 20.4M | 20.4M | 6.9M |
| Training hardware | Not trained by Convergent (Ministral 3 14B) | ASUS TUF RTX 3060 Ti | 1× NVIDIA H100 (Modal) | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti | ASUS TUF RTX 3060 Ti |
Score per category
Every model on this page – our own and the open comparison models – across all Convergent 3.0 categories. Each point is one benchmark score in percent. Best of 10 runs per benchmark.