racineai/Mistral-Military-24B

VISIONPricing:Input $0.867 / Output $1.61Concurrent Unit Cost:2Model Size:24BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 22, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Mistral-Military-24B by Racine.ai is a 24-billion-parameter language model built on Mistral-Small-3.1-24B-Base-2503, specialized for the defense and security domain. It was continued-pretrained on approximately 4 billion tokens of open-source military content across 30 languages, with a focus on French and NATO military doctrine. This model demonstrates a +6.7 percentage point gain over its base model on the Hard Mil Bench, outperforming larger open-source models like Gemma 4 31B and Qwen 3.5 27B in its specialized area. It is intended for research and education in defense policy, strategic studies, and military history.

Loading preview...

Mistral-Military-24B: Domain-Adapted LLM for Defense and Security

Mistral-Military-24B, developed by Racine.ai, is a 24-billion-parameter language model specifically tailored for the defense and security sector. It is based on mistralai/Mistral-Small-3.1-24B-Base-2503 and has undergone continued pre-training on a substantial corpus of open-source military content.

Key Capabilities and Performance

  • Domain Specialization: Focused on French and NATO military doctrine, equipment, and institutional vocabulary, with a multilingual training corpus covering 30 languages (predominantly French and English).
  • Performance Gains: Achieves a +6.7 percentage point improvement over its base model on the Hard Mil Bench, a 561-item open-ended benchmark for military knowledge.
  • Efficiency: Operates at 24 billion parameters, yet matches or outperforms larger open-source models like Mistral-Small 4 (≈119B), Gemma 4 31B, and Qwen 3.5 27B on its specialized benchmark, demonstrating strong performance at a significantly smaller scale.
  • Training Methodology: Utilizes LoRA adapters for continued pre-training on approximately 4 billion tokens, followed by a TIES merge with the instruct variant of the base model to retain conversational abilities while integrating domain knowledge.

Intended Use Cases

This model is designed for research and education only within the defense and security domain. Specific applications include:

  • Research assistance in defense policy, strategic studies, and military history.
  • Educational tool for higher-military education institutions (e.g., IHEDN, École de Guerre).
  • Revision aid for officer trainees and students in military academies.
  • A reference benchmark for research into domain adaptation of large language models.

Finetunes
1 models