Mistral-NeMo-Minitron 8B Instruct
NVIDIA · 8.4B parameters
Mistral-NeMo-Minitron 8B Instruct is NVIDIA's open-weight 8B chat model fine-tuned from Mistral-NeMo-Minitron 8B Base for roleplay, retrieval-augmented generation and function calling, trained in August and September 2024; the BF16 checkpoint contains exactly 8,414,105,600 parameters. It keeps the base model's 40-layer grouped-query-attention decoder and 8,192-token context. The card documents no exact public release date, and the weights are released under the NVIDIA Open Model License.
Benchmark scores
| Benchmark | Score |
|---|---|
| BFCL v2 Live | 67.6 |
| Capability | 111.9 |
| GSM8K | 87.1 |
| HumanEval | 71.3 |
| MBPP | 72.5 |
| MMLU | 70.4 |
| MT-Bench (GPT-4-Turbo judge, corrected references) | 7.9 |