Mistral Large 4: The New Reference Model
On October 6, 2026, Mistral opened Mistral Large 4 to public preview, billed as its largest and most capable model to date. A few weeks later, the company released the weights under the nickname “Le Chonk.” It is now the designated successor to Mistral Medium 3.5 and Mistral Large 3, both deprecated on October 31.
Specifications
| Feature | Value |
|---|---|
| API IDs | mistral-large-4, mistral-large-4-0 |
| Status | Public preview (since October 6, 2026) |
| Architecture | Mixture of experts, 1.05 trillion parameters with 52 billion active |
| Context window | 1 million tokens |
| Inputs / Outputs | Text and image input; text and reasoning output |
| Features | Structured outputs, function calling, document Q&A, batch, agents and conversations, connectors |
| Pricing | $0.70 per million input tokens ($0.07 cached), $2.10 output |
| Weight License | Mistral Ecosystem Open Weight License |
Two lines require close attention. Status first: a public preview can still evolve, which Mistral acknowledges by stating the model continues to improve rapidly. License next: the weights are open, but under a Mistral-specific license, not Apache 2.0 like some older models. Read it before any self-hosted commercial deployment.
What Mistral Highlights
According to the official announcement, Large 4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s data centers in Europe, and the preview is served on the same infrastructure. Mistral positions it as competitive with the best global open source models and significantly superior to any open-weight model developed in the US or Europe, with a particular focus on code, agentic workflows, multimodal understanding, and domains like cybersecurity, finance, and law. These are the vendor’s claims: as always, measure on your own use cases before deciding.
The announcement also emphasizes cybersecurity. Before releasing the weights, Mistral had the model tested in real-world conditions by security professionals, partners, and public authorities. The argument is that a provider’s refusals can block vulnerability research or legitimate incident response, and that open and self-hostable weights give organizations both capability and autonomy.
A First Call
The call doesn’t change compared to other Mistral models; only the identifier evolves.
import os
from mistralai.client import Mistral
client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])
response = client.chat.complete(
model="mistral-large-4",
messages=[{"role": "user", "content": "Summarize a processor's obligations under the GDPR in five points."}],
)
print(response.choices[0].message.content)Should You Migrate Now?
If your code calls Mistral Medium 3.5 or Large 3, you have until April 30, 2027: these models still respond during their deprecation period. Use this time to compare Large 4 on your evaluations, keeping its preview status in mind. For a production service that requires stability, check the official table to see if the model has reached general availability before switching.
Official sources: Introducing Mistral Large 4 · Model Sheet · Changelog
Pricing captured on November 2, 2026 — check the current rate card on Mistral API pricing.
Key Takeaways
- Mistral Large 4 (
mistral-large-4) has been in public preview since October 6, 2026, with open weights published in October - Mixture of experts with 1.05 trillion parameters, 52 billion active, 1 million token context
- $0.70 / $2.10 per million tokens; Mistral Ecosystem Open Weight License for weights
- Designated successor to Medium 3.5 and Large 3, retired on April 30, 2027