Mistral NeMo
Mistral NeMo is a 12-billion-parameter open-weight language model developed jointly by Mistral AI and NVIDIA, notable for its 128K token context window and use of a new, more efficient tokenizer called Tekken.
Definition
Mistral NeMo is a 12-billion-parameter open-weight language model developed jointly by Mistral AI and NVIDIA, notable for its 128K token context window and use of a new, more efficient tokenizer called Tekken.
Overview
Mistral NeMo was released in mid-2024 as a collaboration between Mistral AI and NVIDIA, combining Mistral's model design expertise with NVIDIA's training infrastructure (built and trained on NVIDIA's cloud AI infrastructure using the NeMo framework, from which it takes its name). At 12B parameters it sits between Mistral's smaller 7B models and its larger Mixtral mixture-of-experts models, positioned as a drop-in upgrade for teams already using Mistral 7B who need better reasoning and a much longer context window without moving to a substantially larger, costlier model. A distinguishing feature is its tokenizer, Tekken, built on the tiktoken library and trained on over 100 languages. Mistral reports Tekken is significantly more efficient than the tokenizer used in earlier Mistral models, especially for non-English languages, source code, and compressed text, which reduces token counts (and therefore cost and effective context usage) for the same input. The model supports a 128K token context window, function calling, and was released with both a base and an instruct-tuned checkpoint under the Apache 2.0 license, making it usable commercially without restriction. Mistral NeMo also shipped with an FP8-quantized checkpoint optimized for efficient inference on NVIDIA hardware, reflecting the partnership's focus on production deployability. It is commonly used as a mid-sized, cost-efficient open alternative to closed APIs for multilingual chat, retrieval-augmented generation over long documents, and agentic applications that need both long context and low latency.
Key Features
- 12B parameters, positioned between Mistral 7B and Mixtral
- 128K token context window
- Tekken tokenizer: tiktoken-based, trained on 100+ languages, more token-efficient
- Jointly developed and trained by Mistral AI and NVIDIA using the NeMo framework
- Apache 2.0 license, free for commercial use
- Base and instruction-tuned checkpoints released
- Native function-calling support
- FP8-quantized checkpoint for efficient NVIDIA GPU inference