Mistral Large 4
by mistralai
Mistral's 1T-parameter open-weight model that reads images, charts and documents and answers in text
mistralai/Mistral Large 4Deploy Mistral Large 4
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
Mistral Large 4 is Mistral AI's largest open-weight model, published on Hugging Face on 30 September 2026: a Mixture-of-Experts model with about 1 trillion parameters, 49 billion of them active per token (52 billion counting embeddings and output layers). It takes text and images and returns text, with a 1M-token context, function calling, strict JSON-schema output and a reasoning effort you set per request (low, medium or high).
For a media or document pipeline it is a describer and an extractor. Give it a frame, a chart, a technical drawing or a scanned page and it returns a description, an answer or structured fields. Mistral says its visual grounding leads open-weight models; that is Mistral's claim, and the model card publishes coding and general-intelligence charts rather than vision scores. It returns no retrieval vector, so the text it writes is embedded by a separate embedding model before anything can be searched.
The weights are under the Mistral Ecosystem Open-Weight License, a modified MIT license that allows commercial use and internal use. Offering the model to third parties as a hosted service (Model as a Service) needs a separate commercial license from Mistral.
Architecture
Granular Mixture-of-Experts decoder: 128 routed experts with 4 active per token plus 1 shared expert, about 1T total and 49B active parameters (52B with embeddings and output layers). Pixtral-style vision encoder trained on variable image sizes and aspect ratios. 1M-token context. Ships with an embedded DSpark speculative-decoding draft head that proposes blocks of 3 tokens. Released in full precision and as an NVFP4 checkpoint for Blackwell GPUs; served with vLLM (recommended), SGLang, transformers and llama.cpp.
Mixpeek SDK Integration
# Mistral Large 4 runs on your side (the model card's vLLM example uses
# --tensor-parallel-size 8 on one 8-GPU node) or on Mistral's API. It writes
# a description of each frame; a text embedding model turns that into the
# vector, and both are upserted to a collection.
import base64, requests
from openai import OpenAI
vlm = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="EMPTY")
frame = base64.b64encode(open("ep-412_000184.jpg", "rb").read()).decode()
description = vlm.chat.completions.create(
model="mistralai/Mistral-Large-4-1T-A52B",
reasoning_effort="low",
messages=[{"role": "user", "content": [
{"type": "text", "text": "Describe this video frame in two sentences for a search index: people, objects, on-screen text, setting."},
{"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{frame}"}},
]}],
).choices[0].message.content
requests.post(
"https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
headers={"Authorization": "Bearer API_KEY"},
json={"collection_id": "col_your_collection",
"documents": [{"document_id": "ep-412-000184",
"vectors": {"text-embedding": your_embedding_of(description)},
"payload": {"description": description, "source": "s3://archive/ep-412.mp4", "start_ms": 184000}}]},
)Capabilities
- Image, chart, technical drawing and document understanding with text output
- Visual grounding: pointing to where in an image something is
- Function calling and strict JSON-schema output for structured extraction
- Reasoning effort set per request: low, medium or high (the default)
- 1M-token context and 160+ languages, per Mistral
Use Cases on Mixpeek
Performance
We have not measured it. With 49B active parameters per token, a description costs far more than an embedding, so describe sampled frames or a retrieved short list rather than every frame in an archive.
Common Pipeline Companions
Embeds the descriptions this model writes, which is what makes them searchable
Another 1T-parameter open-weight MoE that takes images, with 32B active parameters
The cheap path when a label or an image embedding is enough and a sentence is more than you need
Frequently Asked Questions
Can I use Mistral Large 4 as an embedding model for image search?
No. It returns text from text and images and has no retrieval vector. Use it to write descriptions or extract fields, then embed that text with an embedding model, or embed the images directly with an image model such as SigLIP 2.
Can I use Mistral Large 4 commercially?
Yes, under the Mistral Ecosystem Open-Weight License, a modified MIT license that covers commercial and internal use. The exception is offering the model to third parties as a hosted service, which needs a separate commercial license from Mistral.
What hardware does Mistral Large 4 need?
The model card's vLLM and SGLang examples use tensor parallelism across 8 GPUs on one node, and Mistral publishes an NVFP4 checkpoint for Blackwell GPUs with a smaller memory footprint. It is also available on Mistral's API.
How does Mistral Large 4 compare with Kimi K2.6?
Both are open-weight Mixture-of-Experts models at about 1 trillion parameters that take images. Mistral Large 4 activates 49B parameters per token and has a 1M-token context; Kimi K2.6 activates 32B with a 256K context. Neither card gives a shared vision benchmark to compare, so test both on your own images.
Specification
Research Paper
Mistral Large 4 technical report
arxiv.orgBuild a pipeline with Mistral Large 4
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free