# Mistral Nemo 12B

A 12B multilingual model with a novel tokenizer trained on 100+ languages, offering ~30% better source code compression, 2x better compression for Korean, and 3x better compression for Arabic than its predecessor.

- **Model ID:** `mistral/mistral-nemo`
- **Type:** chat
- **Providers:** deepinfra
- **Context window:** 131,072
- **Maximum output tokens:** 131,072
- **Pricing:** $0.029/1M input tokens, $0.03/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/mistral-nemo

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'mistral/mistral-nemo',
  prompt: 'Why is the sky blue?'
})
```

## About

Released July 18, 2024, Mistral Nemo 12B was built in collaboration with NVIDIA and introduced the Tekken tokenizer, trained across 100+ languages, as its defining technical innovation. Tekken achieves ~30% better compression for source code compared to previous Mistral tokenizers, 2x better compression for Korean, and 3x better compression for Arabic. These compression gains directly reduce token consumption and cost.

At 12B parameters with a context window of 131.1K tokens, Mistral Nemo 12B serves as a drop-in replacement for Mistral 7B. Mistral Nemo 12B provides enhanced instruction following, multi-turn conversation quality, and code generation. Quantization-aware training enables FP8 inference without performance degradation. The combination of quantization awareness and Tekken compression gives Mistral Nemo 12B deployment efficiency advantages.

Mistral Nemo 12B's multilingual coverage spans English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Korean, Arabic, and Hindi. Mistral Nemo 12B is available under Apache 2.0, with both base and instruct weights on HuggingFace.

## What to consider

The Tekken tokenizer's compression efficiency means that for code-heavy or non-Latin-script workloads, you use fewer tokens per request than with models using conventional tokenizers.

## When to use

### Best For

- Multilingual applications spanning European, East Asian, and Arabic-script languages
- Code-heavy workloads where Tekken's 30% compression advantage reduces token costs
- Korean or Arabic applications where 2-3x tokenizer compression is significant
- Deployments migrating from Mistral 7B that need a larger context window

### Consider Alternatives When

- You need a larger general-purpose model with more headroom (consider Mistral Large 3)
- Your task is primarily code generation or agentic coding (consider Devstral or Codestral)
- Higher reasoning depth is required (consider Magistral models)

## Best for

- **Multilingual applications:** Spanning European, East Asian, and Arabic-script languages
- **Code-heavy workloads:** Tekken's 30% compression advantage reduces token costs
- **Korean or Arabic applications:** 2-3x tokenizer compression is significant
- **Mistral 7B migrations:** Deployments that need a larger context window

## Consider alternatives

- **Larger general-purpose headroom:** You need more capacity (consider Mistral Large 3)
- **Code generation or agentic coding:** Coding is the primary task (consider Devstral or Codestral)
- **Higher reasoning depth:** Tasks that require deeper reasoning traces (consider Magistral models)

## Frequently asked questions

### What is the Tekken tokenizer?

Tekken is a tokenizer trained on 100+ languages, introduced with Mistral Nemo 12B. Tekken achieves ~30% better source code compression, 2x better compression for Korean, and 3x better compression for Arabic compared to previous Mistral tokenizers.

### What is the context window for Mistral Nemo 12B?

131.1K tokens.

### Is Mistral Nemo 12B a drop-in replacement for Mistral 7B?

Yes. Mistral positions it as a drop-in upgrade with the same architecture family, improved quality, and a larger context window.

### What languages does Mistral Nemo 12B support?

English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Korean, Arabic, and Hindi, among others.

### What is FP8 inference and how does quantization-aware training help?

FP8 is a reduced-precision number format that speeds inference and reduces memory usage. Quantization-aware training means the model was trained to tolerate FP8 quantization, so accuracy doesn't degrade compared to full-precision inference.

### What is the license for Mistral Nemo 12B?

Apache 2.0, permitting commercial use and modification.

### Who built Mistral Nemo 12B?

Mistral in collaboration with NVIDIA, as indicated by the NeMo branding aligned with NVIDIA's NeMo framework ecosystem.

## Links

- [Model page](https://vercel.com/ai-gateway/models/mistral-nemo)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://deepinfra.com/mistralai/Mistral-Nemo-Instruct-2407)
- [Provider pricing](https://deepinfra.com/pricing)