Mistral NeMo

Claims and evidence

• Released 18 July 2024 ● Primary source[1]page date; first paragraph [4]July 18, 2024 [5]release date [2]first paragraph
• Deprecated 22 May 2026 ● Primary source[6]Deprecated & retired models table, row Mistral Nemo 12B [5]Deprecation date
• Retired 31 July 2026 ● Primary source[6]Deprecated & retired models table, row Mistral Nemo 12B

• Status AvailableApache 2.0 weights remain downloadable from the mistralai Hugging Face organisation; the API endpoint open-mistral-nemo-2407 was deprecated on 2026-05-22 and retired on 2026-07-31. ● Primary source[7] [3] [6]Deprecated & retired models table, row Mistral Nemo 12B
• Replaced by Ministral 3Mistral names Ministral 3 8B, a variant of the Ministral 3 record. ● Primary source[6]Deprecated & retired models table, row Mistral Nemo 12B, Alternative [5]Replacement

• Successor of Mistral 7B v0.3 EditorialMistral presents Mistral NeMo as a drop-in replacement for Mistral 7B and compares the two, without naming a Mistral 7B version; linking to v0.3, the Mistral 7B version current on la Plateforme at the time, is an editorial choice. ● Primary source[1]first paragraph; Instruction fine-tuning [3]Key features

• Change · Size Has 12B parameters, up from 7B in Mistral 7B, while keeping a standard architecture that Mistral says lets it replace Mistral 7B as a drop-in.Compared with Mistral 7B v0.3 ● Primary source[1]
• Change · Architecture Uses a new tokenizer, Tekken, based on Tiktoken, in place of the SentencePiece tokenizer of earlier Mistral models.Compared with Mistral 7B v0.3 ● Primary source[1]
• Change · Efficiency Trained with quantisation awareness to allow FP8 inference, which Mistral says comes without a loss in performance.Compared with Mistral 7B v0.3 ● Primary source[1]
• Change · Languages Designed for multilingual use; Mistral names eleven languages, among them Chinese, Japanese, Korean, Arabic and Hindi.Compared with Mistral 7B v0.3 ● Primary source[1]

• Input text ● Primary source[5]Modalities
• Output text ● Primary source[5]Modalities
• Feature Function calling ● Primary source[1]Multilingual Model for the Masses [5]Features
• Feature Multilingual ● Primary source[1]Multilingual Model for the Masses
• Open weights YesBase and instruct checkpoints under Apache 2.0. ● Primary source[1]second paragraph; Links [7] [3]
• Context window 128k tokens ● Primary source[1]first paragraph [5]Context
• Parameters 12B ● Primary source[1]first paragraph [5]Weights table
• Access API, Open-weights download ● Primary source[1]Links

Lineage

Variants

No variants recorded in this record.

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Mistral AI on AI Radar →