Mistral Nemo Instruct 2407
A 12B model that Mistral AI trained jointly with NVIDIA, instruction-tuned for chat in many languages and for function calling. The card calls it a drop-in replacement for Mistral 7B.
Overview
Mistral Nemo Instruct 2407 is the instruction-tuned checkpoint of Mistral-Nemo-Base-2407, a 12-billion-parameter model that Mistral AI built together with NVIDIA and released in July 2024. A large share of its training data was multilingual text and code. Mistral AI calls it a drop-in replacement for Mistral 7B, and its release post says Nemo is much better than that model at following precise instructions and handling multi-turn conversations.
The network is a 40-layer transformer with grouped-query attention (32 query heads, 8 key-value heads). Its tokenizer, Tekken, is built on Tiktoken and was trained on more than 100 languages; the vocabulary holds about 128k tokens. Against the SentencePiece tokenizer of earlier Mistral models, Mistral AI reports roughly 30% better compression for source code, Chinese and several European languages, 2x for Korean and 3x for Arabic.
What you can build
- Customer support in the user's own language. Mistral AI names English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Korean, Arabic and Hindi as its strongest.
- Agents that call your functions for bookings or stock checks, then write the reply from what comes back.
- Explanations and reviews of short pieces of code.
- Story drafting, as in the sample, which turns Tower of London history into five titled pitches.
Inputs
A plain prompt works for one-off questions. For multi-turn work you can pass the full conversation as a messages list instead, leaving the prompt empty. Wiro can also keep the history for you when you set a user_id and session_id.
Functions are attached as OpenAI-format definitions. Tool choice accepts auto, none, required or one specific function. Of these, only none changes what the model does, by keeping the functions out of that turn. The Parallel Tool Calls checkbox (on by default) does not limit a turn either, so the model decides how many calls to return. A second run, carrying the first run's task token, returns the function results to the model.
Below these sit the sampling controls. The one worth changing is temperature (see Recommended settings).
Outputs
Text, or one or more function calls with JSON arguments. The sample answers the Tower of London prompt with five events, each followed by a titled story idea such as "The Stonemasons' Tale".
Recommended settings
- Lower the temperature to about 0.3. Mistral AI's card says Nemo needs smaller temperatures than earlier Mistral models and recommends 0.3, while this page starts at 0.7.
- Set tool choice to none when you want a plain reply even though functions are attached. Required or a named function does not force a call here, so check that one came back before acting on it.
Limitations
- Mistral AI trained Nemo with a 128k context window, but a request on Wiro holds about 35,800 input tokens. In a long conversation the oldest turns go first.
- Check facts before you use them. In the sample, Guy Fawkes is found guarding the gunpowder in the Tower's cellars; the barrels were actually under the House of Lords.
- Quality varies by language. The card lists 68.0% on MMLU (5-shot), and its MMLU scores in other languages run from 59.0% for Chinese and Japanese to 64.6% for Spanish.
Safety & compliance
Nemo Instruct comes with no moderation mechanisms, according to Mistral AI's card, so content filtering is up to your application. Mistral AI released both the base and the instruct checkpoints under Apache 2.0.
API quick start
Run Mistral-Nemo-Instruct-2407 with a single API call.
{
"prompt": "What are some interesting historical even…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}curl -X POST "https://api.wiro.ai/v1/Run/mistralai/mistral-nemo-instruct-2407" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "What are some interesting historical even…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}
JSONOther models you might like
Explore similar llm & chat models.



