Mistral 7B Instruct v0.3
Mistral AI's 7B chat model, fine-tuned for instructions from Mistral-7B-v0.3. This version added function calling and moved to the v3 tokenizer with a 32,768-token vocabulary.
Overview
Mistral 7B Instruct v0.3 is Mistral AI's instruction fine-tune of Mistral-7B-v0.3, a base model with about 7.25 billion parameters. Compared with v0.2, the vocabulary grew to 32,768 tokens and the tokenizer moved to v3. Function calling is also new in this version. Mistral AI presents the instruct model as a demonstration of how easily the base model can be fine-tuned, and publishes the weights in BF16.
Give each request a user_id and a session_id and Wiro stores the chat history under them, so one conversation can carry on over many requests.
What you can build
- A support or FAQ bot whose scope and tone are fixed by the system prompt.
- Short drafting jobs such as emails and product blurbs.
- Answers to questions about a passage you paste into the prompt.
- A small agent that picks one of a few functions, for example an order-status lookup, and fills in its arguments.
- Tagging and extraction jobs that need a short reply in a fixed format.
Inputs
Each run takes either a prompt or a messages list holding the whole conversation, never both. The prefilled prompt asks about historical events near the Tower of London and how they could inspire a story. The system prompt comes with a general helpful-assistant instruction that you can replace.
Wiro admits about 17,400 input tokens per request. A chat that grows past this loses its earliest turns, and a long document should be sent in parts.
Temperature runs from 0 to 5 and starts at 0.7. Top-p, top-k, repetition and length penalties, token limits, semicolon-separated stop sequences and a seed (123456 by default) are optional.
Function definitions go in the tools field in OpenAI format. Results from your functions are sent back with the previous task's token in the next run.
Outputs
Plain text, or a function call with a name and JSON arguments when definitions are attached. The sample on this page answers the default prompt with five numbered events, from the building of the White Tower to the Gunpowder Plot, each with a one-sentence story idea. One entry shows how loosely the model can treat facts: it lists the murder of Thomas Becket as an event near the Tower while saying that it happened in Canterbury Cathedral.
Recommended settings
- Turn sampling off for extraction and labeling, so the same input returns the same answer.
- Leave max tokens at 0 and the reply length is sized automatically. Set a number when you need a hard cap.
Limitations
Function calls are not guaranteed. The model sometimes answers in prose where a call was expected, so check that a call came back and validate its arguments before acting on it.
Safety & compliance
The weights are released under Apache 2.0. Mistral AI's card says the model does not have any moderation mechanisms, and that the team hopes to work with the community on making it respect guardrails. If your product needs moderated output, filter prompts and replies in your own application.
API quick start
Run Mistral-7B-Instruct-v0.3 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-7b-instruct-v0-3" \
-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.
