Mistral Small 4

mistral-small-4-119b 🇨🇭 Switzerland

Mistral Small 4 di Mistral AI (marzo 2026): legge immagini, usa strumenti e ragiona solo quando una richiesta lo chiede (reasoning_effort "high"). Una mistura di esperti da 119 miliardi di parametri, circa 6,5 per token, quindi risponde in fretta. Contesto di 262.144 token, pesi a 4 bit (NVFP4) pubblicati da Mistral AI.

At a glance

Publisher	Mistral AI
Architecture	Mixture of experts, 119 billion parameters, about 6.5 billion per token
Licence	Apache 2.0
Precision	NVFP4, 4 bits (Mistral)
Context	262,144 tokens, prompt and answer together
Longest answer	`max_tokens` up to 8,192, reasoning included
Reads	Text and images
Calls tools	Yes
JSON schema	Yes
Reasons	Yes, before every answer. Switch it off with reasoning_effort: "none"; the reasoning counts as output tokens.
Runs in	🇨🇭 Switzerland
Status	Loaded: the first request answers without a start-up wait. Live state on Status.
Price	CHF 0,15 input · CHF 0,60 output, per million tokens, excluding VAT. All prices

Use it

curl https://api.ai-tokens.it/v1/chat/completions \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-small-4-119b",
    "reasoning_effort": "none",
    "messages": [{"role": "user", "content": "Summarise in one line: the meeting moves to Thursday at 10."}]
  }'

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.ai-tokens.it/v1", api_key=os.environ["API_KEY"])
answer = client.chat.completions.create(
    model="mistral-small-4-119b",
    messages=[{"role": "user", "content": "Summarise in one line: the meeting moves to Thursday at 10."}],
)
print(answer.choices[0].message.content)

More in Quickstart and Chat completions.

Limits

Prompt and answer together stay within 262,144 tokens; beyond that the engine refuses the request with a 400 that says so.
The reasoning counts against max_tokens: a low cap can end the answer before it starts, with finish_reason: "length".
Served at 4 bits, which makes it faster and fits our machines; we write the precision of every model.
A request goes only to the zones your key allows; if the model has no machine there now, a 503 names the zones. It is never sent elsewhere.