Mistral NeMo

What it is

Mistral NeMo is a 12-billion-parameter language model released on 18 July 2024, built by Mistral AI in collaboration with NVIDIA^[1]. It supports a context window of up to 128K tokens and was positioned as a drop-in replacement for Mistral 7B^[1]. Base and instruction-tuned checkpoints were released under the Apache 2.0 licence^[1].

The context window is the amount of text a model can keep in view at once, counting the prompt and the answer together; a 128K window means a single request can carry a long document, a large source file or a long conversation rather than a short excerpt. "Drop-in replacement" means a system already built around Mistral 7B could switch to the new model without redesigning how it is called.

How it was released

The model came out of a collaboration between Mistral AI and NVIDIA^[1]. Its weights were published on Hugging Face, and the same model was offered on La Plateforme under the name open-mistral-nemo-2407 and packaged as an NVIDIA NIM microservice^[1]. Shipping both a base checkpoint and an instruction-tuned one under a permissive licence gave users two starting points: a raw model to fine-tune further, and a model already tuned to follow instructions. The model card describes Mistral-Nemo-Instruct-2407 as an instruct fine-tune of Mistral-Nemo-Base-2407^[3].

The Tekken tokenizer

The release introduced a new tokenizer called Tekken^[1]. A tokenizer is the component that cuts text into the units, called tokens, that a model actually reads and generates; how much text fits in a context window depends directly on how those cuts are made. Tekken is based on Tiktoken and was trained on more than 100 languages^[1]. On source code and several major languages it is about 30% more efficient, twice as efficient on Korean and three times as efficient on Arabic^[1]. Efficiency here means fewer tokens for the same piece of text, which matters because both the context window and the compute budget are measured in tokens: the same file or conversation takes up less of the window, and costs less to process.

Training choices and language coverage

The model was trained with quantisation awareness, which makes FP8 inference possible without a loss of performance^[1]. Quantisation stores a model's weights in fewer bits so that it runs faster and takes less memory; normally this trades away some accuracy, and training with the reduced precision in mind is meant to avoid that trade. Mistral NeMo covers English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Korean, Arabic and Hindi, and it was trained for function calling^[1] — the ability to emit a structured call to an external tool instead of an answer written in prose.

Benchmarks and comparisons

At release Mistral compared the model against Gemma 2 9B and Llama 3 8B^[1]. The model card reports 68.0% on MMLU with 5-shot prompting and 73.8% on TriviaQA with 5-shot prompting for the instruct model^[3], and 83.5% on HellaSwag and 76.8% on Winogrande, both measured with 0-shot prompting^[3]. MMLU is a multiple-choice test drawn from many school and university subjects; TriviaQA is a question-answering set; HellaSwag and Winogrande test how well a model picks a sensible continuation or resolves which entity a phrase refers to. "5-shot" and "0-shot" describe how many worked examples are placed in the prompt before the question. Scores like these are comparable only when the prompting setup matches, which is why the number of shots is recorded alongside each figure.

Where it sits

The model marks a step in the line of open-weight Mistral releases: a mid-sized model with a long context window, a retrained tokenizer aimed at code and non-English text, and a distribution arrangement in which the same weights arrived through Hugging Face, a hosted API and an NVIDIA service^[1]. Mistral NeMo was deprecated on 22 May 2026 and retired on 31 July 2026, replaced by Ministral 3 8B^[2]. That end date is part of the record: the release stayed in service for about two years before being superseded by a newer model.