Mistral Large 4: the Open-Weight Preview You Can’t Download Yet

1. What Mistral shipped, and what is still a promise

Mistral Large 4 is a granular Mixture-of-Experts model with 1.05T total and 52B active parameters plus a 1.6B vision encoder, according to Mistral’s model card (version v26.10, status Public Preview, dated October 6). Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and that the preview is served on that same infrastructure.

The word to watch is “open-weight”. Mistral’s post says it “will release the weights by the end of the month” and the model card tags the model as Open, but today nothing is downloadable and no license has been named. Treat everything about self-hosting as a plan, not a fact.

2. What improves over Mistral Large 3

Mistral’s launch post does not compare Large 4 with Large 3, so I set the two spec sheets side by side. Large 3 (Hugging Face repo Mistral-Large-3-675B-Instruct-2512, published November 2025) is a ~675B-total / ~41B-active MoE with a 256K context window, released under Apache 2.0 according to its card.

Large 4 grows on three axes: roughly 1.5x the total parameters, about 27% more active parameters and, on paper, four times the context. It also folds reasoning into the same model. It is not cheaper: the example on Mistral’s pricing page still prices Mistral Large at $0.50 / $1.50 per 1M tokens, so even the $0.68 / $2.09 sale price is about 1.4x that, and list price about 2.7x.

The three real differences

3. Evidence: vendor numbers and independent ones

Most numbers in Mistral’s launch post are Mistral’s own, and the post does not say which reasoning setting produced them. Read them as direction, not as a ranking.

The independent signals are thinner. Artificial Analysis scores the preview 38 on its Intelligence Index (rank 65 of 216 on its page, class median 27) and reports about 200M output tokens to run the index against a median of 85M: this is a verbose model. Simon Willison, who tried the API, puts it just behind DeepSeek V4.1 Flash on that index and says it is not Fable-class.

4. Quickstart from Mistral’s own docs

The preview speaks the same chat-completions API as every other Mistral model. Install the official SDK (mistralai on PyPI, 3.2.0 released October 9, 2026, Python 3.10+; the JavaScript package @mistralai/mistralai is at 2.7.0) and create an API key in Mistral Studio.

Before you wire it into anything, do two things: list the models your key can see, and read the context window the API actually reports. The model ID appears as mistral-large-4-0 on the model card and in the reasoning guide, and as mistral-large-4 in the changelog, so let the API settle it.

# pip install mistralai
import os
from mistralai.client import Mistral

with Mistral(api_key=os.environ["MISTRAL_API_KEY"]) as mistral:
    res = mistral.models.list()
    for m in res.data:
        if m.id.startswith("mistral-large-4"):
            print(m.id, m.aliases, m.max_context_length)

import os
from mistralai.client import Mistral
from mistralai.client.models import TextChunk, ThinkChunk

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])

resp = client.chat.complete(
    model="mistral-large-4-0",   # confirm with models.list()
    messages=[{"role": "user", "content": "Find the race condition in this handler: ..."}],
    reasoning_effort="high",     # documented values: "high" | "none"
)

content = resp.choices[0].message.content
if isinstance(content, str):     # "none" returns a plain string
    print(content)
else:                            # "high" returns thinking + text chunks
    for chunk in content:
        if isinstance(chunk, TextChunk):
            print(chunk.text)    # the final answer
        elif isinstance(chunk, ThinkChunk):
            pass                 # reasoning trace: keep in history, do not show

print(resp.usage)                # see how many tokens "high" really cost

5. Honest limits

Large 4 is a preview with open-weight intentions. These are the limits I would put in front of a team before they pick it.

6. What to do before the weights ship

Treat the next two weeks as a free evaluation window, not as a migration. Nothing below needs the weights.

Verdict

Mistral Large 4 is a credible, EU-hosted, trillion-class model with a clear open-weight promise and a coding profile that Mistral’s own numbers place below the best closed models. As of October 14 it is a preview you can test, not an open model you can run.

Evaluate it through the API this week with your own prompts and token accounting, keep your current model as the fallback, and come back when the weights and a named license appear. If they do not land by the end of October, plan as if it were a hosted model.