Mistral NeMo (12B)

// Overview

Mistral NeMo is a 12-billion-parameter open dense language model that Mistral AI built jointly with NVIDIA and released on July 18, 2024. It ships under the permissive Apache 2.0 license, with both a base checkpoint (Mistral-Nemo-Base-2407) and an instruction-tuned checkpoint (Mistral-Nemo-Instruct-2407) on Hugging Face. The model offers a 128K-token context window and was positioned by Mistral as a state-of-the-art replacement for the older Mistral 7B in its size class.

A distinguishing feature of Mistral NeMo is its new tokenizer, called Tekken, which is based on Tiktoken and trained on over 100 languages. Mistral reports that Tekken compresses natural-language text and source code more efficiently than the SentencePiece tokenizer used in earlier Mistral models, with especially large gains for languages such as Korean and Arabic. The model is multilingual by design, with Mistral highlighting English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Korean, Arabic, and Hindi.

Mistral NeMo was trained with quantization awareness so it can run FP8 inference without losing accuracy, and it was designed to fit on a single GPU, making it practical for local and self-hosted deployment. On Mistral's own platform the model was served as open-mistral-nemo-2407. Mistral has since marked that API endpoint as deprecated (deprecation date 5/22/2026, retirement 7/31/2026) and points API users to Ministral 3 8B as the recommended successor, but the open weights remain freely available under Apache 2.0 and are widely mirrored through tools like Ollama and LM Studio.

// Benchmarks

1. MMLU (5-shot)68%
2. HellaSwag (0-shot)83.5%
3. Winogrande (0-shot)76.8%
4. OpenBookQA (0-shot)60.6%
5. CommonSenseQA (0-shot)70.4%
6. TruthfulQA (0-shot)50.3%
7. TriviaQA (5-shot)73.8%
8. NaturalQuestions (5-shot)31.2%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

First-party Mistral la Plateforme pricing for open-mistral-nemo. The API endpoint is deprecated (retirement 7/31/2026); open weights are free to self-host under Apache 2.0. Third-party hosts may price differently.

Pricing source ↗

// Strengths

• Apache 2.0 open weights — base and instruct checkpoints are freely usable, modifiable, and self-hostable
• Large 128K-token context window, trained natively rather than extended after the fact
• Efficient Tekken tokenizer covering 100+ languages, with strong multilingual and code compression
• Quantization-aware training enables FP8 inference with no reported accuracy loss
• Compact 12B size fits on a single GPU, making it cheap to run locally
• Drop-in replacement for Mistral 7B thanks to a standard transformer architecture

// Best for

• Self-hosted chat and assistant applications on a single GPU
• Multilingual text generation and translation across 100+ languages
• Long-document summarization and analysis using the 128K context
• Function calling and tool use in lightweight agent workflows
• On-premises or privacy-sensitive deployments where open weights are required
• Code generation and completion for everyday programming tasks

// How to access

// Mistral 7B / Nemo (open dense) — every version

The full lineage of the Mistral 7B / Nemo (open dense) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.