NewsAI & Development

Mistral Large 4: 1T Open Weights Drop — Free to Use

Conceptual illustration of Mistral Large 4 open-weight AI model release with 1 trillion parameter visual

Mistral AI dropped the open weights for Large 4 today. A 1 trillion-parameter model, free to download, fine-tune, and deploy commercially — no gating, no waitlist. Nicknamed “Le Chonk” with full self-awareness, it is the largest open-weight model to come out of a Western AI lab, and the weights are live on Hugging Face right now.

The Security Angle Nobody Is Talking About

The most interesting part of this release is not the benchmark numbers. It is what the model will do that its competitors will not.

Claude Opus 5.5 and GPT-6 Astra score near zero on vulnerability reproduction benchmarks — not because they lack capability, but because their safety filters refuse the work entirely. Ask either one to reproduce a real CVE for patch verification and you will hit a wall. Mistral Large 4 scores 82% on vulnerability reproduction and patching, solved 93% of Cybench’s CTF challenges, and hit 79% on CyberGym-E2E, which tests whether a model can find, reproduce, and patch memory-safety bugs in real C and C++ codebases.

This was a deliberate product decision. Security teams doing legitimate pen testing, CVE triage, and vulnerability research have been squeezed out by safety filters calibrated for a different risk profile. Mistral made a bet that serving those teams is worth the tradeoff — and given that AI-assisted exploit research is already happening regardless of guardrails, it is a defensible one.

Strong Numbers, Honest Context

Large 4 sits second globally on the Coding Agent Index at 49.8%, just behind GPT-6 Astra. On DeepSWE it hits 62%, ahead of DeepSeek-V4-Pro at 57% and GLM-5.3 at 61%. Terminal-Bench 4.0, which tests autonomous coding in a real terminal environment, comes in at 41%. Notably, it achieves this while using 47% fewer tokens per task than GLM-5.3, which translates to roughly 63% lower cost-per-task on agentic workflows.

However, the honest comparison is this: Large 4 scores 38 on Artificial Analysis’s Intelligence Index. Claude Opus 5.5 scores 58. The gap is real. Mistral reached where OpenAI was seven months ago. For teams that need open weights for compliance, data sovereignty, or fine-tuning freedom, Large 4 is the best option outside China right now. For teams that simply want the strongest model on a given task, it is not.

Can You Actually Run This Locally?

“Open weights” does not mean “runs on your laptop.” Here is the hardware reality:

QuantizationDownload SizeMin RAMWho Can Run It
2-bit (Q2)~400 GB512 GBMac Studio M5 Ultra (barely)
4-bit (Q4)~650 GB768 GBTwo clustered Macs or a server
8-bit (Q8)~1.1 TB1.2 TB+Server-class only

The minimum viable local setup is a 512 GB Mac Studio M5 Ultra running the 2-bit quantized version, with careful memory allocation adjustments just to clear the threshold. The 49 billion active parameters per token create real memory bandwidth pressure — do not expect fast inference even on top-tier consumer hardware. For most developers, the practical path is the API: $0.68 per million input tokens, $2.10 per million output.

The License in Plain English

The Mistral Ecosystem Open-Weight License is a modified MIT license. You can use it, modify it, fine-tune it, build products with it, and sell those products. The one catch: if you are operating a Model-as-a-Service business that charges customers specifically for API access to this model, you need a commercial license from Mistral. Building an app that uses Large 4 under the hood? You are fine. Wrapping the weights and selling inference as a service to others? Talk to Mistral first. Full license details are here.

How to Get It

Weights are live on Hugging Face under mistralai/Mistral-Large-4-1T-A52B-NVFP4. Community GGUF builds from Unsloth are also available for llama.cpp users. vLLM and SGLang are the recommended serving frameworks for self-hosters. Managed options include Mistral Studio, NVIDIA Build, and Azure AI Foundry.

The European angle is built in: trained on 3,800 Grace Blackwell GPUs across European data centers, covering 160+ languages, and fully operable on-premises under EU law. Per Fortune’s coverage, Mistral frames this directly at EU banks, governments, and regulated enterprises that need frontier-class AI with no foreign dependency. The Decoder’s analysis puts it plainly: a European bank can now download this, run it on-premises, and use it for cybersecurity without asking anyone’s permission. That is the actual product.

ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *

    More in:News