News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnMistral NeMo

Mistral NeMo

A 12B multilingual open-weights model built with NVIDIA, with a 128k context window.

Mistral NeMo is a 12 billion parameter language model released by Mistral AI together with NVIDIA. It ships as base and instruction-tuned checkpoints under Apache 2.0, handles a 128k token context window, and targets multilingual use and code.

Share on WhatsApp

  • 12B parameters, built jointly by Mistral AI and NVIDIA
  • 128k token context window
  • Base and instruct weights released under Apache 2.0
  • Tekken tokenizer trained on over 100 languages
  • Trained with quantisation awareness for FP8 inference

What it is

Mistral NeMo is a 12 billion parameter large language model put out by Mistral AI in partnership with NVIDIA. It comes in two checkpoints: a pre-trained base model and an instruction-tuned version for chat and task following. The company positions it as a strong option within its size class for reasoning, world knowledge and coding, and as a straightforward swap for systems already running Mistral 7B. It was trained with quantisation awareness, so it can run FP8 inference without reported loss of quality.

How it came about

Mistral AI announced the model as a joint build with NVIDIA. Both the base and instruction-tuned weights were released at the same time under Apache 2.0, aimed at researchers and enterprises. The instruction-tuned checkpoint is described by its creators as a quick demonstration that the base model fine-tunes easily to strong performance. On Mistral's platform it is exposed under the name open-mistral-nemo-2407.

Inside the model

The model is a standard transformer, which Mistral says makes it easy to deploy. It uses the SwiGLU activation function, grouped-query attention with 8 key-value heads, rotary position embeddings with theta set to 1M, and a vocabulary of about 128,000 tokens. It was trained on a large share of multilingual and code data with a 128k token context window. A new tokenizer called Tekken, built on Tiktoken and trained on more than 100 languages, replaces the SentencePiece tokenizer used by earlier Mistral models. An advanced fine-tuning and alignment stage followed pre-training.

What it can do

Mistral AI reports the base model as competitive with Gemma 2 9B and Llama 3 8B, and says it leads its size category on reasoning, world knowledge and coding accuracy. The model card lists zero-shot scores of 76.8% on Winogrande, 60.6% on OpenBookQA, 70.4% on CommonSenseQA, 50.3% on TruthfulQA, and 31.2% on NaturalQuestions at five shots. Instruction tuning is said to improve precise instruction following, reasoning, multi-turn dialogue and code generation over Mistral 7B. It is trained for function calling.

How to use it

Base and instruct weights are published on Hugging Face under Apache 2.0. They can be run with mistral-inference, Hugging Face transformers or NVIDIA's NeMo, and adapted with mistral-finetune. Mistral also serves the model on la Plateforme as open-mistral-nemo-2407, and NVIDIA packages it as a NIM inference microservice available from ai.nvidia.com. The company recommends a temperature of about 0.3, lower than for earlier Mistral models.

Safety and testing

The model card states plainly that the instruct model has no moderation mechanisms built in. Its authors say they want to work with the community on guardrails so the model can be deployed where moderated output is required.

Why it matters

Mistral AI frames the release as a step toward making capable models available across the languages of human culture. Tekken is presented as a lasting contribution: about 30% better compression than SentencePiece for source code and several European languages, twice as good for Korean and three times for Arabic, and better than the Llama 3 tokenizer for roughly 85% of languages tested.

Questions people ask about Mistral NeMo

What is Mistral NeMo?

It is a 12 billion parameter language model from Mistral AI, developed with NVIDIA. It ships as a pre-trained base checkpoint and an instruction-tuned version, both released under the Apache 2.0 licence, and is aimed at multilingual work and code.

How does it differ from Mistral 7B?

Mistral AI says NeMo is a drop-in replacement for Mistral 7B in existing systems. It follows precise instructions better, reasons better, handles multi-turn conversation better and generates stronger code, and it uses the new Tekken tokenizer instead of SentencePiece.

What is Tekken?

Tekken is the tokenizer Mistral NeMo uses, built on Tiktoken and trained on more than 100 languages. It compresses text and code better than the earlier SentencePiece tokenizer, with the largest gains on Korean, Arabic and several European languages.

How can I run it?

Weights are on Hugging Face and work with mistral-inference, transformers or NVIDIA NeMo. Mistral also offers it through la Plateforme as open-mistral-nemo-2407, and NVIDIA ships it as a NIM container from ai.nvidia.com.

Does it have content moderation?

No. The model card says the instruct version has no moderation mechanisms, and that its makers hope to work with the community on guardrails for deployments that need moderated outputs.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Mistral NeMo yet, so the facts on this page come from the sources listed above: mistral.ai/news/mistral-nemo, huggingface.co/mistralai/Mistral-Nemo-Instruct-2407. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.