TokenstromAll about AI Toolbox

Mistral Large 4 "Le Chonk": a 1-trillion-parameter model with open weights

Mistral has released the weights of Mistral Large 4, a model with roughly one trillion parameters that understands text and images. It leads open models in coding but trails the best Chinese open-weight models overall. Self-hosting requires server hardware.

Model releases Try it yourself Oct 29, 2026 5 min read Source: Mistral AI
Illustrative image, AI-generated

What it's about

Paris-based Mistral AI has released the weights of Mistral Large 4, known internally as "Le Chonk." The model debuted as a public preview three weeks earlier. "Weights" are the numerical values a model learns during training. Whoever has them can run and customize the model themselves. The model accepts text and images and outputs text and code. Mistral stresses that it was trained from scratch entirely in Europe, and markets it as a building block for "sovereign AI."

The weights are on Hugging Face. You can also access the model through Mistral Studio, the Mistral API, OpenRouter, Microsoft Foundry and NVIDIA endpoints.

Architecture and tech

According to the Hugging Face blog, Mistral Large 4 is a mixture-of-experts (MoE) model. The network consists of many specialized sub-networks ("experts"), of which only a few are active per token. In total it has about 1 trillion parameters (roughly, the knobs the model has learned). According to Hugging Face, 49B are active per token, while The Decoder says 52B. The repo name "A52B" supports the 52B figure, so the sources contradict each other.

Details from the blog:

  • 72 layers, hidden dimension 6,144, 64 attention heads.
  • Multi-head Latent Attention (MLA) with a compressed cache, combined with a DeepSeek Sparse Attention indexer that selects up to 2,048 earlier positions per query.
  • The first layer is dense, and the remaining 71 are MoE blocks. A router picks 4 of 128 experts, and a shared expert runs as well, so there are 5 active experts per token.
  • A 1.6B-parameter Pixtral image encoder processes images.
  • The context window is 1M tokens. According to Mistral, the model supports 160+ languages.

There are three reasoning levels (Low, Medium, High; High is the default). Reasoning means the model "thinks" before answering. According to Hugging Face, it can't be turned off. "none" still uses High.

Coding performance

After three more weeks of reinforcement learning (training by rewarding correctly solved tasks), Mistral reports significant progress. The figures are mostly vendor claims:

Benchmark Result
Terminal-Bench 4.0 41% (Preview: 28%), GLM-5.3: 40%
Artificial Analysis Coding Agent Index 48% (virtually tied with GLM-5.3, but 47% fewer tokens)
FrontierCode 2.0 (Cognition) 19.7% (Claude Opus 5.5: 46%)
BrowseComp 84.5%
Finch (spreadsheets/finance) 73.3%
B3 Agent Security 97.1% attacks repelled

In blind comparisons by Surge Code, testers preferred Large 4 more often than GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Against Claude Opus 5.5, it won only 22 percent of comparisons, according to The Decoder. The Hugging Face blog gives different values (for example 28.3% on Terminal-Bench 4 and 93.3% on B3), probably from the preview phase.

Cybersecurity

Mistral wants to position the model for defenders: finding, reproducing and patching vulnerabilities. On the Artificial Analysis Cyber Index it scores 51 percent, according to Mistral. The Decoder notes that in Mistral's own chart, DeepSeek V4.1 Flash (57%) and MiMo-V2.6-Pro (56%) lead, which makes Large 4 only third among open models. The preview scored 50 points, so the gain is small. The Hugging Face blog cites 82 percent for the "reproduce and patch" sub-test and 93 percent on Cybench.

According to Mistral, the model refuses malicious cyber requests 96.6 percent of the time on average. Mistral also admits the model could make real attacks easier, and plans to tighten the safeguards in the coming months. Partners such as Orange Cyberdefense are testing a specialized variant for penetration testing, with vetting and contractual conditions.

Pricing, license, availability

The API costs $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens. In the preview it was $1.36 and $4.18. The license allows use, customization, fine-tuning and commercial use. Only those who offer the model as a cloud service (model-as-a-service) need a separate license.

Companies can tailor the model to their own data with Mistral Forge. Mistral says Large 4 will serve as the basis for further specialized models. The company didn't comment on smaller variants.

Independent measurements

Artificial Analysis has released version 5 of its Intelligence Index. Large 4 (High) scores 26 points, the preview 25. DeepSeek V4.1 Flash and GLM-5.3 score 32, MiMo-V2.6-Pro 30, and the median of similarly sized open models is 23. The leaders are Gemini 4 Carbon (59, currently only for selected cyber partners) and Claude Opus 5.5 (58). The best open score is Kimi K3.1 Pro with 46, but its weights aren't due until November 4.

Its strengths are speed and cost: 161 tokens per second (Preview: 111), and an index task costs $2.06 on average (Preview: $5.54). That's cheaper than GLM-5.3 ($5.01) but more expensive than DeepSeek V4.1 Flash ($0.60). Token usage fell from 150 to 110 million, but is still slightly above the median of 100 million.

What you need to try it

The model is built for data centers. According to the SGLang documentation, the standard version runs on eight Nvidia B300 GPUs, and a more heavily compressed variant on four GB300. The NVFP4 checkpoint in vLLM requires Blackwell GPUs. A normal home GPU won't cut it. For llama.cpp, there are GGUF quantizations (such as Q2_K) from Unsloth and the llama.cpp team. Even these require a huge amount of memory, and the source gives no exact figures. For simple testing, the API, OpenRouter or Inference Providers are more practical.

Our take

An open model of this size from Europe is notable, especially for organizations that care about data sovereignty. It leads open models in coding, but is far behind Claude Opus 5.5 and will likely be overtaken soon by Kimi K3.1 Pro. On the overall index, it stays behind Chinese models. The "lead" in cyber defense is put in perspective by The Decoder's counter-analysis. Many figures are vendor claims, and the sources partly contradict each other on active parameters and benchmark values.

Sources

#mistral#open-weights#coding#cybersecurity#moe
Read the original source

Try it yourself: Mistral Large 4

Explained step by step, no prior knowledge needed – on your own computer or free in your browser.

What can I do with it?

Mistral Large 4 is a huge model that understands text and images and is meant to shine especially at programming, agentic tasks and IT security. It can write code, fix bugs in projects, reproduce and patch vulnerabilities, edit spreadsheets and research the web. It speaks over 160 languages and has a very large context window of 1M tokens. According to Mistral, it leads open models in coding, but overall it trails Chinese open-weight models like GLM-5.3 and DeepSeek V4.1 Flash and falls well behind Claude Opus 5.5. Locally it only runs on server hardware, so for ordinary users the API is the realistic route.

Who is it worth it for?

Interesting for companies, research teams and developers with access to GPU clusters or an API who value European origin and self-hosting. The model isn't suitable for ordinary PCs and laptops.

Ideas: What you can do with it

Coding agent in your own repo

The model gets a bug or a feature in your own repository, for example via Vibe Code. It writes changes and tests them itself.

Vulnerability triage

A security team has it review reports, reproduce flaws in a sandbox test and create patches. The result is verifiable fixes.

Editing financial spreadsheets

The model creates and edits spreadsheets for accounting and controlling. You describe the structure you want and get a finished file.

Image analysis

You upload photos or screenshots, and the model answers questions about them or extracts information.

Custom specialist models

Researchers use the large model's answers to generate training data for fine-tuning smaller security models.

Size1000 B parameters (49 billion active)
Download635 GB (about 14.1 hr at 100 Mbit/s)
Licensecustom license (see Model Card)
TypeImage + text (understands images)
Hugging Facemistralai/Mistral-Large-4-1T-A52B

What your computer needs

Graphics card (NVIDIA), fully in VRAM at least 420 GB VRAM, recommended 700 GBservers only, with multiple data-center GPUs (e.g. 8× H100) – fastest; with less VRAM, see the next line
Graphics card + system memory (MoE offloading) from approx. 21 GB VRAM + 671 GB RAMdie Experten liegen im Arbeitsspeicher, nur der Kern auf der Grafikkarte (llama.cpp/LM Studio). Langsamer als komplett im VRAM, aber auf normalen PCs machbar
Mac (Apple Silicon M1–M5) too big for any MacThe GPU and CPU share the memory – the model can use about 70% of it
CPU only (no graphics card) Very slow without a graphics cardan estimated 1 words per second on a current desktop PC (MoE: only 49 billion parameters active per word). This requires at least 671 GB of RAM.
Storage approx. 637 GB freeModel download plus programs
Google Colab too big for Colabeven the largest Colab GPUs aren't enough

Which variant is right for me?

Many models come in several sizes or "quantizations." These are compressed versions: smaller and faster, but slightly less accurate. Rule of thumb: pick the largest variant whose memory requirement fits in your VRAM (video memory). The scripts below choose this for you automatically.

VariantDownloadNeeds approx.Compatible graphics cards
Q2_K 381 GB420 GBservers only, with multiple data-center GPUs (e.g. 8× H100)
Q4_K_M Default 635 GB700 GBservers only, with multiple data-center GPUs (e.g. 8× H100)
Q8_0 1115 GB1228 GBservers only, with multiple data-center GPUs (e.g. 8× H100)
BF16 2099 GB2310 GBservers only, with multiple data-center GPUs (e.g. 8× H100)

Officially there are the standard FP8 checkpoint and an NVFP4 variant (smaller, Blackwell GPUs only). According to SGLang, the standard version runs on eight B300 GPUs and the compressed one on four GB300. GGUF quants such as Q2_K are available for llama.cpp, but still need a lot of memory.

  • Standard version: eight Nvidia B300 according to the SGLang docs. Compressed version: four GB300.
  • The NVFP4 checkpoint in vLLM requires Nvidia Blackwell GPUs.
  • In the Hugging Face example, vLLM is launched with tensor-parallel-size 8 and max-model-len 131072.
  • A consumer GPU or a normal laptop is not enough.
  • GGUF quants (e.g. Q2_K) run with llama.cpp, but still need an enormous amount of RAM/VRAM.

Groß (1000 B parameters, MoE): running entirely on the GPU takes a lot of VRAM (video memory). With MoE offloading, a regular graphics card plus plenty of RAM is enough (see table). Colab Free isn't enough.

How do I find out what my computer has?
  • Windows: Press Ctrl + Shift + Esc → Task Manager → Performance tab → click GPU on the left. At the bottom you'll see Dedicated GPU memory (that's your VRAM). You'll find system memory under Memory.
  • Mac: Apple menu at the top left → About This Mac. You'll see your chip (e.g. M2) and memory (e.g. 16 GB). On Apple chips, graphics and processor share this memory.
  • Linux: Type nvidia-smi in the terminal (shows your graphics card and its memory) and free -h for system memory.

Guide

Google Colab (in your browser, no installation)

Google Colab is a free service from Google. It gives you a computer with a graphics card in your browser for a few hours. All you need is a Google account. For this model: too big for Colab – even the largest Colab GPUs aren't enough.

  1. Download the notebook. Click the button and save the file.

    Download colab.ipynb

  2. Open Colab. Go to colab.research.google.com and sign in with your Google account.
  3. Upload the notebook. In the window that appears, click Upload on the left (or choose File → Upload notebook at the top) and select the file colab.ipynb.
  4. Turn on the GPU. In the top menu, go to Runtime → Change runtime type → under Hardware accelerator, select T4 GPU (with Colab Pro: L4 or A100) → Save.
  5. Start everything. Runtime → Run all (keyboard shortcut Ctrl + F9). If you see the warning “This notebook was not authored by Google,” click Run anyway.
  6. Wait. The cells run one after another. A green checkmark appears to the left of each cell when it's done. Installation and download take about 216 minutes here.
  7. Try it out. The result appears below the last cell. To test something else, change the text in the cell and run it again with Shift + Enter. The last cell starts a proper chat with an input field.
  8. Clean up. When you're done: Runtime → Disconnect and delete runtime. This saves your free quota.

Good to know: The free GPU is limited (usually a few hours per day). After about 90 minutes of inactivity, Colab disconnects, and you'll have to start everything over.

Windows 10 / 11

The easiest way is Ollama, a free program that launches AI models with a single command.

  1. Update your graphics driver (NVIDIA cards only): Install the NVIDIA app or the latest driver. For AMD cards, use the latest Adrenalin driver.
  2. Install Ollama. Download the file OllamaSetup.exe from ollama.com/download/windows, double-click it, and choose Install. Afterward, a small llama icon appears in the taskbar at the bottom right.
  3. Open a terminal window. Press the Windows key, type PowerShell and press Enter. A blue or black window opens.
  4. Start the model. Copy this command ("Copy" button), paste it into the window with a right-click, and press Enter: 
    ollama run hf.co/ggml-org/Mistral-Large-4-1T-A52B-GGUF:Q4_K_M
    The first time, Ollama downloads the model (635 GB (about 14.1 hr at 100 Mbit/s)). After that, it's instant.
  5. Start typing. Once >>> Send a message appears, type your question and press Enter. Type /bye to end the chat.

Not enough memory? Pick a smaller variant from the table above, e.g. ollama run hf.co/ggml-org/Mistral-Large-4-1T-A52B-GGUF:Q2_K.

Prefer automatic? A script that picks the right variant for you
  1. Download both files into the same folder: try_windows.ps1 demo.py
  2. Open the folder in File Explorer, right-click an empty spot and choose Open in Terminal.
  3. Paste this command and press Enter:
    powershell -ExecutionPolicy Bypass -File try_windows.ps1

If needed, the script installs Ollama via winget, measures your VRAM (video memory) and downloads the matching variant.

macOS

Macs with an Apple chip (M1 to M5) are ideal. Older Intel Macs are very slow.

  1. Install Ollama. Download it from ollama.com/download/mac, open the file, and drag Ollama into your Applications folder. Launch Ollama once – a llama appears in the menu bar at the top.
  2. Open Terminal. Press Cmd + Space, type Terminal, and press Enter.
  3. Start the model. Paste the command (Cmd + V) and press Enter: 
    ollama run hf.co/ggml-org/Mistral-Large-4-1T-A52B-GGUF:Q4_K_M
    The first time, the model is downloaded (635 GB (about 14.1 hr at 100 Mbit/s)).
  4. Start typing. Type your question and press Enter. Quit with /bye.
Pick the right variant automatically

try_mac.sh

cd ~/Downloads
bash try_mac.sh

Linux

  1. Check your GPU. nvidia-smi should show your card. If it doesn't, install the driver on Ubuntu and restart: 
    sudo ubuntu-drivers install
  2. Install Ollama: 
    curl -fsSL https://ollama.com/install.sh | sh
  3. Start the model: 
    ollama run hf.co/ggml-org/Mistral-Large-4-1T-A52B-GGUF:Q4_K_M
    Exit with /bye. Use ollama ps to check whether the model is running on the GPU.
  4. Or let it pick the right variant automatically:

    try_linux.sh

    bash try_linux.sh
# mit Python/transformers statt Ollama:
MODE=hf bash try_linux.sh

Try it

Who are you and what can you do?
How many cats can you see in this picture? (with image)
Write a Python script that reads a CSV file and plots monthly revenue as a chart.
Analyze this code for security vulnerabilities, explain them and suggest patches.
Create a spreadsheet for household bookkeeping with formulas for totals and month-over-month comparison.

Tips for better results

  • Use Medium for everyday tasks and High only for tricky problems. It saves tokens and costs.
  • For a first test, the API or OpenRouter is worth trying before you build out hardware.
  • In the llama.cpp WebUI (localhost:9931), you can add images and set the reasoning level.

If something goes wrong

Update transformers first: pip install --upgrade transformers.

According to Hugging Face, reasoning can't be turned off. "none" still uses High.

The NVFP4 variant only runs on Blackwell GPUs.

Model-as-a-service providers need a separate commercial license.

For Transformers, the script must be launched with torchrun on multiple GPUs.

“out of memory,” a crash, or your computer freezes

The model is too big for your VRAM (video memory). Pick a smaller variant from the table above and close other programs, especially games and browsers with lots of tabs.

It runs, but extremely slowly

Your CPU is probably doing the work instead of your graphics card. In Ollama, the PROCESSOR column of ollama ps shows how much runs on the GPU. If it says "100% CPU", update your graphics driver or pick a smaller variant that fits entirely into VRAM (video memory).

“ollama” or “python” is not found

Close the terminal window and open it again so it recognizes the newly installed program. For Python on Windows: reinstall it and check Add python.exe to PATH.

Windows: "Running scripts is disabled on this system"

Run the script with exactly the command above: powershell -ExecutionPolicy Bypass -File try_windows.ps1. This allows execution for this one time only.

Error message when loading the model (version, "trust_remote_code", unknown architecture)

Brand-new models often need the latest software version. With Ollama: update Ollama. With Python: pip install -U transformers in the activated environment. The Model Card usually lists the exact version.

Tokenstrom generates the scripts and notebook automatically. Please take a quick look before running them – they're plain, readable text. Mind the model's license: custom license (see Model Card).

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.

More articles

Model releases Kakao open-sources two Kanana-2 embedding models, built for Korean search Model releases Clover-35B: Diffusion language model claims twice the agent speed of Jev Model releases Krea 2 fine-tune places comics and text exactly with coordinate boxes