On this page
Mistral AI is a French company, founded in 2023, that releases most of its language models as open weights — you can run them locally, fine-tune them, or call them through the hosted API. For developers who want capable European models with permissive open-source licensing, it is one of the best free options in 2026.
Free Models on La Plateforme
These are the same open-weight models published on Hugging Face, but here you call them through a hosted endpoint with no infrastructure to manage.
| Model ID | Size | Context | Best For |
|---|---|---|---|
open-mistral-nemo | 12B | 128k tokens | General, multilingual, summarization |
open-mistral-7b | 7B | 32k tokens | Fast responses, prototyping |
open-mixtral-8x7b | 56B MoE | 32k tokens | 429 on a free key (Sept 2026) |
open-mixtral-8x22b | 141B MoE | 64k tokens | 429 on a free key (Sept 2026) |
codestral-latest | 22B | 32k tokens | Code generation and completion |
Mistral Nemo is the recommended starting point — trained with NVIDIA, 128k context, native English/French/Spanish/German. For code, codestral-latest is a strong Copilot alternative. Free-tier access is rate-limited; paid tiers unlock higher throughput plus Mistral Small and Large.
Checked September 29, 2026 on a free key: open-mistral-nemo, open-mistral-7b, codestral-latest and the ministral-3b/8b/14b-latest models all answered. open-mixtral-8x7b and open-mixtral-8x22b returned 429, and so did mistral-small-latest and every mistral-medium model — even with 12 seconds between calls. The response header explains it: x-ratelimit-limit-req-minute: 0 for mistral-small-latest, against 188 for ministral-8b-latest. A limit of zero means the free plan simply doesn't include those models; waiting won't help. mistral-large-latest returned 403 "not available in your subscription tier".
How fast is it? We measured
We ran Mistral through our open benchmark against every other free AI API — same US machine, same prompt, timed from each API's own token counts (2026-08-31):
| Provider | First token | Generation |
|---|---|---|
| Groq | 0.38 s | 527 tok/s |
Mistral (mistral-small-latest, 31 Aug — no longer callable on our free key) | 0.43 s | 167 tok/s |
| Google Gemini | 34.5 s | 141 tok/s |
| NVIDIA NIM | 3.5 s | 33 tok/s |
Two honest readings of that. On 31 August Mistral was second fastest overall and its 0.43-second first token was essentially tied with Groq's — unlike Gemini's 34-second pause. But on raw generation Groq was three times quicker, so if throughput is the only thing you care about, Mistral isn't the pick. Caveat: those numbers are for mistral-small-latest, which a free key can no longer call. On 2026-10-09 the models a free key can call took 0.69 s (ministral-8b-latest) to 4.98 s (codestral-latest) to the first token; we have not re-measured their generation speed.
Where Mistral earns its place is the combination: a genuinely free tier with no credit card, European hosting, and models tuned for multilingual work. That mix is rarer than raw speed. Full comparison: free AI APIs benchmarked.
Get a Free Key and Make Your First Call
Sign up at console.mistral.ai (no credit card), verify your email, then API Keys → Create new key and copy it — it's shown only once. Set it as export MISTRAL_API_KEY="your_key_here", install the SDK, and call any free model:
pip install mistralai
import os
from mistralai import Mistral
client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])
response = client.chat.complete(
model="open-mistral-nemo",
messages=[
{"role": "user", "content": "Explain how mixture-of-experts models work in plain English."}
]
)
print(response.choices[0].message.content)
For real-time output, stream instead:
with client.chat.stream(
model="open-mistral-nemo",
messages=[
{"role": "user", "content": "Write a step-by-step guide to building a REST API with FastAPI."}
]
) as stream:
for text in stream.get_text_stream():
print(text, end="", flush=True)
Node.js is the same shape (npm install @mistralai/mistralai):
import Mistral from "@mistralai/mistralai";
const client = new Mistral({ apiKey: process.env.MISTRAL_API_KEY });
const response = await client.chat.complete({
model: "open-mistral-nemo",
messages: [
{ role: "user", content: "What are the key differences between Mistral 7B and Mixtral 8x7B?" }
]
});
console.log(response.choices[0].message.content);
Drop-In Replacement for the OpenAI SDK
Mistral's API is fully OpenAI-compatible. If you already call OpenAI, switch in two lines — no new dependencies, no restructuring:
from openai import OpenAI
# Change these two lines to switch from OpenAI to Mistral
client = OpenAI(
api_key=os.environ["MISTRAL_API_KEY"],
base_url="https://api.mistral.ai/v1"
)
response = client.chat.completions.create(
model="open-mistral-nemo",
messages=[
{"role": "user", "content": "Summarize the main benefits of using Mistral AI."}
]
)
print(response.choices[0].message.content)
Code Generation with Codestral
Codestral supports fill-in-the-middle (FIM) completion — give it the start and end of a function and it fills the middle. It handles Python, JavaScript, TypeScript, Java, Go, Rust, SQL, and 80+ languages:
from mistralai import Mistral
client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])
# Fill-in-the-middle code completion
response = client.fim.complete(
model="codestral-latest",
prompt="def calculate_fibonacci(n):n ",
suffix="n return result"
)
print(response.choices[0].message.content)
Mistral vs Other Free AI APIs
| Provider | Free Tier | Best Free Model | Speed | Strengths |
|---|---|---|---|---|
| Mistral AI | Rate-limited free models | Mistral Nemo 12B | Fast | Open-source, EU privacy, code model |
| Google Gemini | Free tier; caps not published | Gemini Flash | Very fast | 1M context, multimodal |
| Groq | 30 RPM, 1,000 req/day (chat models) | gpt-oss-120b | Fastest (LPU) | Speed, low latency |
| DeepSeek | None — top-up required | deepseek-v4-flash | Fast | Cheap paid reasoning |
| GitHub Models (retired) | — | — | — | Shut down 2026-07-30 |
| Cloudflare Workers AI | 10K neurons/day | Llama 3.1 8B | Edge-fast | Global edge, no cold start |
Where Mistral stands out: it's the only major provider here that publishes open-source weights under a permissive license. Prototype via API, then self-host the same Hugging Face models for cost savings — no lock-in.
Pricing (When You Need More)
| Model | Input (per 1M) | Output (per 1M) | Context |
|---|---|---|---|
| Mistral Small 3.1 | $0.10 | $0.30 | 128k |
| Mistral Large 2 | $2.00 | $6.00 | 128k |
| Codestral | $0.30 | $0.90 | 32k |
| Mistral Embed | $0.10 | — | 8k |
When you hit the free rate limits, Mistral Small at $0.10/1M input tokens is one of the cheapest production-grade models available — versus $0.15/1M for GPT-4o Mini.
When to Use Mistral AI
Reach for Mistral when you need a European provider for GDPR or data residency, want to prototype via API then self-host the same weights, need a strong free code model (Codestral), want OpenAI-compatible drop-in with minimal code changes, or need multilingual support beyond English.
Consider alternatives when you need a huge context window (Gemini Flash, 1M tokens), the fastest inference (Groq), or frontier reasoning that requires GPT-4o or Claude.
Final Verdict
Mistral fills a unique niche: open-source weights, an EU data center, and a fully OpenAI-compatible API. The drop-in compatibility is the practical killer feature — switching existing OpenAI code takes about 30 seconds. Grab a free key at console.mistral.ai and your first call is two minutes away.