Mistral NeMo 12B
Mistral NeMo 12B was trained jointly by Mistral AI and NVIDIA. It provides base and instruction-tuned weights under Apache 2.0 with an advertised 128K context.
Deprecated API; check access details
A different tokenizer and training mix
The release introduced Tekken, a tokenizer intended to encode multilingual text and code efficiently. Mistral also describes quantization-aware training for FP8 inference. An FP8 package changes representation and serving cost; it is not another NeMo model family.Mistral NeMo 12B release announcement
Choose Base or Instruct
Mistral-Nemo-Base-2407 is the pretrained checkpoint. Mistral-Nemo-Instruct-2407 adds instruction fine-tuning. The architecture has 40 layers, grouped-query attention and a vocabulary of 131,072 entries. Use the corresponding tokenizer and prompt format for the selected variant.Mistral NeMo 12B official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Deprecated API; check access details |
| Launch date | |
| Access | API, Downloadable |
| Model identifiers | Mistral-Nemo-Base-2407Mistral-Nemo-Instruct-2407open-mistral-nemo-2407 |
| Context window | 128,000 tokens The model catalog advertises 128k context for this version. This is not a measured recall limit. Mistral NeMo 12B model and lifecycle record |
| Weight license | Apache 2.0Mistral NeMo 12B official model card |
| Access and lifecycle | Hosted API deprecated on 22 May 2026. Downloadable weights remain listed.Mistral NeMo 12B model and lifecycle recordMistral NeMo 12B official model card |
| Co-developers | Mistral AI and NVIDIAMistral NeMo 12B official model card |
| Parameters | 12 billionMistral NeMo 12B model and lifecycle record |