abcl-servicing-intent · Models
Which model should Dvar use to score your traffic?
The judge, topic names, rubric drafts and alternative replies call these models with your tested key, through the Dvar connector.
Models on Dvar Inference
Open-weight models served from India. Tested ones show their share of benchmark messages fully right; the judge that scores replies is anthropic/claude-sonnet-5.5.
| Model | Price per 1M tokens (in / out) | On this benchmark |
|---|---|---|
| DeepSeek V4 Flash dvar:deepseek-v4-flash-0731 | ₹10 / ₹20 | 91.9% fully right |
| Gemma 4 26B dvar:gemma-4-26b-a4b-it | ₹10 / ₹30 | 95.9% fully right |
| Gemma 4 31B dvar:gemma-4-31b-it | ₹10 / ₹30 | wrote the synthetic messages |
| GLM 5.3 dvar:glm-5.3 | ₹105 / ₹340 | 96.3% fully right |
| GLM 5.3 Flash dvar:glm-5.3-flash | ₹10 / ₹40 | 92.3% fully right |
| gpt-oss-120b dvar:gpt-oss-120b | ₹10 / ₹40 | 97.2% fully right |
| Muse Glimmer 30B dvar:muse-glimmer-30b | ₹25 / ₹110 | 86.6% fully right |
| Nemotron 3.5 30B dvar:nemotron-3.5-lightning-30b-a3b-bf16 | ₹5 / ₹20 | 94.4% fully right |
| Qwen 3.8 27B dvar:qwen-3.8-27b-fp8 | ₹25 / ₹210 | not tested |