Mistral: Mistral Nemo

Mistral Nemo is a 12B-parameter language model built by Mistral AI together with NVIDIA. Its instruct version, Mistral-Nemo-Instruct-2407, has a 128k-token context window and is released under the Apache 2.0 license. On the Mistral API it was served as open-mistral-nemo-2407, and Mistral's models documentation lists it as deprecated on May 22, 2026 and retired on July 31, 2026, with Ministral 3 8B as the replacement.

Status

The hosted API endpoint is retired. The open weights on Hugging Face remain published under Apache 2.0 for self-hosting.

Key Features

• Size and context: 12B parameters, 128k-token context window.
• Tokenizer: Tekken.
• Architecture details from the card: 40 layers, 5,120 dimensions, 32 attention heads with 8 KV heads, and a vocabulary of about 128k.
• Multilingual and code data: the card says a large share of training data is multilingual and code.
• Function calling: supported through transformers 4.42.0 or later. Tool call IDs must be exactly 9 alphanumeric characters.
• Vendor benchmarks: according to the model card, MMLU (5-shot) 68.0%, HellaSwag (0-shot) 83.5%, and TriviaQA (5-shot) 73.8%, with multilingual MMLU between roughly 59% and 65%.
• Recommended temperature: 0.3, per the card.

Use Cases

• Self-hosted chat and function-calling workloads that need a permissive license.
• Fine-tuning experiments on a 12B base.

Limitation: the card says the model has no moderation mechanisms, so add your own safety layer.

Migration Steps

1. Search for open-mistral-nemo-2407 in your code.
2. Move hosted calls to Ministral 3 8B or another model from Mistral's models page, and re-run your evaluations.
3. If you need Nemo specifically, self-host the open weights from Hugging Face.

FAQ

Is Mistral Nemo still available on the Mistral API?

No. Mistral lists it as retired on July 31, 2026.

What replaced Mistral Nemo?

Mistral's documentation names Ministral 3 8B as the replacement. See also Mistral Small in this catalog.

What is Mistral Nemo's context window and license?

128k tokens, under the Apache 2.0 license.

Who built it?

Mistral AI and NVIDIA.

Alternatives

• Mistral Small: a current Mistral model line in this catalog.
• Mistral Large 4: Mistral's larger flagship entry.
• Llama 3.1 8B Instruct: a small open-weights alternative from Meta.

Conclusion

Mistral Nemo is retired on the hosted API but remains a usable open-weights model. See the model card and the models category.