Skip to content
NewSQL, Python and Power BI training for your team
Log in

Mistral Large 4 Is Out: Europe’s Trillion-Parameter Open AI Model

You can now download Le Chonk. Here is what it means, what it costs to run, and how to try it from Nigeria.

Muhammad Jamilu Abdullahi 6 min read
Cover: a glowing sphere of parameter dots with a few lit amber nodes, beside the headline "Mistral Large 4" and a download window showing 610 GB

A giant you can download

Picture a developer in Yaba, Lagos, refreshing a Hugging Face page on Thursday, 29 October. That day, French AI company Mistral posted the weights of Mistral Large 4, nicknamed "Le Chonk". Anyone could now download a model with roughly a trillion parameters.

An earlier Hugging Face listing had pointed to 31 October, so the files came two days early. Then our developer hits the catch. The smallest official version is about 610GB.

That is the whole story in miniature. A frontier-class model is now yours to keep, but running it is another matter.

What Mistral released

Mistral first opened Large 4 as a public preview on 6 October, through Mistral Studio and its API. It then spent about three more weeks training it with reinforcement learning, focused on coding and cybersecurity.

SpecMistral Large 4
Total parametersAbout 1.05 trillion
Active per token49bn (52bn counting embedding and output layers)
DesignMixture of experts: 128 experts, 4 picked per token, plus 1 shared
Vision1.6bn-parameter image encoder; text and images in, text out
Context windowUp to 1m tokens
Languages160+
TrainingAbout 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European data centres
Download sizesAbout 1.06TB (FP8); about 610GB (4-bit NVFP4)
LicenceMistral Ecosystem Open-Weight License (a modified MIT licence)
API price$0.70 per million input tokens; $2.10 per million output tokens

Mistral pitches the model at coding, cybersecurity, finance and AI-agent work. Security firms including Booz Allen, CrowdStrike and Elastic, plus the EU cyber agency ENISA, have been testing it on real security tasks.

Parameters, experts and open weights, in plain English

Parameters are the numbers a model adjusts while it learns. Think of them as billions of tiny dials that store patterns of language and reasoning. More dials usually means more ability, and more memory to hold them.

A "mixture of experts" splits those dials into specialist groups. For each token, a chunk of a word, a router wakes a few experts and leaves the rest idle. Large 4 holds about a trillion parameters but uses roughly 49bn per token, which makes it faster and cheaper to run.

"Open weights" means you can download those learned numbers and run the model on machines you choose. Closed models, such as OpenAI's GPT models or Google's Gemini, are rented through apps and APIs, and their weights stay on company servers. Open weights is not full open source, though, because Mistral has not published its training data.

Why it matters for Nigerian builders

The first win is control. Mistral says organisations can inspect, customise and deploy Large 4 on their own infrastructure, without relying on its API. Pierre Stock, Mistral's VP of Science, told TechCrunch that open-weight access "can't be cut off by definition."

That matters under the Nigeria Data Protection Act 2023. Section 41 bars sending personal data out of Nigeria unless the recipient offers adequate protection or another listed condition applies. Fines for data controllers of major importance reach ₦10m or 2% of annual gross revenue, whichever is higher.

Keeping a model on servers you control, in a location you choose, can make that compliance story simpler. Treat this as general guidance, not legal advice. Banks, hospitals and fintechs should still consult a data-protection lawyer.

The licence is friendly, with one hook. You can use, modify, fine-tune and sell products built on Large 4, and internal use is exempt. But a company selling hosted access to the model itself, as a "Model-as-a-Service", needs a separate commercial licence from Mistral.

The catch: heavy hardware, mixed reviews

The hardware bill is brutal. The 4-bit checkpoint is about 610GB before any working memory, and the FP8 version is about 1.06TB. Coverage of Mistral's guidance says the 4-bit build needs four of NVIDIA's B200 or B300 data-centre GPUs.

For scale, a top-end RTX 5090 gaming card has 32GB of memory. You would need the memory of about 19 of them just to hold the smallest file.

Benchmarks are the second catch. Mistral says the final model scores 41% on Terminal-Bench 4.0, up from 28% at preview, and calls it the best open-weight coding model. Those are Mistral's own figures, and no independent lab had reproduced the 41% by 30 October.

Independent testers were cooler on the preview. The Register reported that Artificial Analysis ranked it between DeepSeek V4.1 Flash and OpenAI's entry-level GPT6 Luna, well behind US flagships. On Vals.ai, despite the finance and legal pitch, it ranked 18th of 34 in finance and 25th of 31 in law.

It is not even the biggest open model. Moonshot released weights for its 2.8-trillion-parameter Kimi K3 in July. Alibaba followed in August with Qwen3.8, at 2.4 trillion parameters.

What you can do now, and what to watch

Start with the API. The model ID is mistral-large-4, at $0.70 per million input tokens and $2.10 per million output tokens. LLM Rumors priced a busy test month, with 100m input, 40m cached and 20m output tokens, at about $115.

Mistral's free Experiment plan needs only a verified phone number and no card. Check which models it covers before you build on it. Keep customer data out too, because Mistral says Experiment requests may be used to train its models.

OpenRouter lists Large 4 at the same price with a 1m-token context, so one key lets you compare it with other models. Mistral also names NVIDIA-hosted endpoints and Microsoft Foundry as routes in.

Want something private on your own laptop today? Mistral's Ministral 3 models, at 3bn, 8bn and 14bn parameters, carry the permissive Apache 2.0 licence. Mistral says they have optimised deployments for RTX PCs and laptops.

Next, watch for independent tests of the final build, Mistral's promised technical report and a vetted cybersecurity variant now with Orange Cyberdefense. At launch, co-founder Guillaume Lample warned that with a closed model there is "no guarantee it will still be there tomorrow." For Nigerian teams, the smart move is to learn on APIs and small models now, then own more of the stack as hardware gets cheaper.

Sources

Keep reading

All posts