The model announcement is out. The API can be tried. For your own local installation, however, the actual package is still missing. This in-between phase is particularly easy to overlook with open models: a technical target picture quickly sounds like a download that is already available.
Mistral presented Large 4 on 6 October 2026 as an API public preview. The announcement gives the end of the month for the release of the weights; the model card likewise lists a public preview. As of the sources on 25 October, preview access and the announced local model package remain separate states.
What the service takes care of
With an API, your application sends a request to a service that provides the model. That service maintains the necessary execution environment. For a small test, you therefore do not first have to fit all of the model weights onto your computer.
A local route shifts this responsibility. You need the right files, a supporting runtime and sufficient resources. On top of that come the applicable terms of use. Having a file does not yet answer every one of these questions; an announcement does not even answer the question of which concrete artefact exists.
For technical planning, it helps to separate three areas: access to the service, complete artefacts and a demonstrated way of running the model. For each area, note the status you have actually checked. That way, a later download becomes a targeted addition rather than a silent correction of an availability claim made too early.
The task stays independent of the access route
First define a small use case, such as summarising your own product notes. Three factual sentences, no added unknown properties and a visible note on contradictions are useful requirements. These rules apply both to a test via the service and to a possible later local comparison.
Use a few synthetic notes whose meaning you know. One names a material, one corrects a quantity, one is missing the size. Keep the inputs and the desired limits unchanged. If both the access route and the instruction change later, you can hardly attribute differences to a single cause.
Identifying an artefact unambiguously
As soon as you actually have a model file that you are permitted to use, you can record its name, its size and a checksum. The following Python helper, written by us, reads an existing file block by block. It downloads nothing, changes no file and does not check model quality.
from pathlib import Path
import hashlib
import sys
pfad = Path(sys.argv[1])
if not pfad.is_file():
raise SystemExit("Specify an existing file.")
pruefer = hashlib.sha256()
with pfad.open("rb") as quelle:
for block in iter(lambda: quelle.read(1024 * 1024), b""):
pruefer.update(block)
print("File:", pfad.name)
print("Bytes:", pfad.stat().st_size)
print("SHA256:", pruefer.hexdigest())Save it as datei-stand.py and pass the actual path of an existing file when you run it. For a model that consists of several parts, the checksum of a single part is not enough for the whole package. In that case, list the relevant artefacts individually and keep a note of their origin with them.
A checksum is a mark of identity. It shows whether two file versions match. It certifies neither a right of use nor security, factual correctness or sufficient memory. These questions belong in separate checks.
From download to documented local operation
For genuine local proof, you also need the exact runtime, its configuration and a small task that has actually been run. A successful download alone is not yet a successful load. A loaded model alone is not yet a usable result.
So keep the command you actually used and the output you observed together with the artefact version. If a prerequisite is missing, exactly that step remains open. Technically, that is more helpful than the blanket statement that the model is now available locally.
Mistral’s news opens a concrete API preview route and describes a planned release of the weights. Anyone who keeps these levels apart can follow the development with interest and later add a traceable test of their own. The clean boundary between announced, present and executed makes a big model announcement understandable in practice.
Sources and status
Sources as of 25 October 2026. Introducing Mistral Large 4 · Mistral Large 4 model card.
