Back to Gap Map Model components / Fine-tuned / chat models

Mistral NeMo

Mistral AI
open weights / Overall score: 3.0

Mistral NeMo is a 12B-parameter language model that Mistral AI built with NVIDIA, released in July 2024 as base and instruct checkpoints with a 128K-token context window. It introduced Mistral's Tekken tokenizer, was trained for function calling and multilingual use, and was designed as a drop-in replacement for Mistral 7B.

Mistral retired the hosted open-mistral-nemo API on July 31, 2026 and names Ministral 3 8B as its replacement. The open weights remain on Hugging Face, and this entry scores them.

Openness

3 high confidence
3.0
weights
open (base and instruct checkpoints, ungated on Hugging Face)
data
closed
code
closed (no training or post-training pipeline published)
license
Apache-2.0 (OSI)

Mistral NeMo's base and instruct weights are published under Apache-2.0 without a gate. The training data and the instruction-tuning recipe are not disclosed, so it is open weights rather than open source.

Adoption

3 high confidence
3.0

Hugging Face downloads summed across the base and instruct checkpoints.

Capability

3 medium confidence
3.0

Mistral NeMo scores 68 on MMLU and was released as a stronger drop-in replacement for Mistral 7B, measured by Mistral against Gemma 2 9B and Llama 3 8B rather than against 7B models, which puts it a step above Mistral 7B Instruct. As a 2024 12B model it is well short of current mid-size releases.

  • https://docs.mistral.ai/getting-started/models/models_overview/ recorded 2026-11-12

    Mistral's models page lists "Mistral Nemo 12B" (open-mistral-nemo-2407) with deprecation 5/22/2026, retirement 7/31/2026 and Ministral 3 8B as the replacement.

  • https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407/raw/main/README.md recorded 2026-11-12

    Card "Main Benchmarks" table: "MMLU (5-shot) | 68.0%", "HellaSwag (0-shot) | 83.5%", "Winogrande (0-shot) | 76.8%", "TriviaQA (5-shot) | 73.8%"; and "Drop-in replacement of Mistral 7B".

  • https://mistral.ai/news/mistral-nemo recorded 2026-11-12

    Release post dated July 18, 2024: "a 12B model built in collaboration with NVIDIA"; a table "compares the accuracy of the Mistral NeMo base model with two recent open-source pre-trained models, Gemma 2 9B, and Llama 3 8B"; and "Compared to Mistral 7B, it is much better at following precise instructions, reasoning, handling multi-turn conversations, and generating code."

Verified 2026-11-12