Mistral Releases Large 4 Weights and Halves the API Price. What About Arabic?

On 29 Oct 2026, Mistral released the open weights for Mistral Large 4, nicknamed Le Chonk, following an API preview running since 6 Oct 2026. As detailed in Mistral’s announcement, three weeks of extra reinforcement learning improved coding and defensive cybersecurity while using fewer tokens to reach answers. The model was trained from scratch in Mistral’s own European data centres, and the API price drops by roughly half at the same time. Our take: it is a serious option for companies that want to host a frontier-class model themselves, but nothing an individual can run at home.

What Changed

Mistral Large 4 converts the earlier preview into a downloadable Mixture-of-Experts architecture. It totals about 1 trillion parameters (1.05 trillion in documentation) with 52 billion active parameters (49 billion per token excluding embeddings and output layers), plus a 1.6 billion vision encoder that takes text and images in to output text. It features a 1M token context window and three reasoning modes: low, medium, and high (high is default), with low replacing the old none mode.

Weights use the Mistral Ecosystem Open-Weight License, a modified MIT licence allowing free modification, fine-tuning, and commercial use; only Model-as-a-Service providers need a separate commercial licence. Regarding safety, Mistral states the model refuses most malicious cyber requests, though it is adjusting guardrails after seeing possible attack uplift in live traffic, reserving a specialized cyber variant for vetted partners like Orange Cyberdefense.

Who It Affects

This release targets enterprise engineering teams and cybersecurity operators who need infrastructure independence. As reported by Wired, Mistral pitches the model as a Western open-weight alternative that cannot be “switched off remotely”, though Wired points out open weights can also be downloaded by attackers. If you already use Mistral’s products (see our Mistral AI listing), this is now the company’s largest model, and you can download it and inspect it yourself.

However, individual developers cannot host this locally. Self-hosting requires data-centre hardware: SGLang documentation states the standard release requires eight Nvidia B300 GPUs, a compressed variant needs four GB300 GPUs, and Hugging Face deploys ready endpoints on 8x H200 setups. Readers following our Ollama vs LM Studio comparison should note this cannot run on laptops.

What It Costs

Alongside the weights release on 29 Oct, Mistral roughly halved its API prices from the preview rates of $1.36 per million input tokens and $4.18 per million output tokens:

• Input tokens: $0.70 per million
• Cached input tokens: $0.07 per million
• Output tokens: $2.10 per million

One caveat on “halved”: The Decoder notes that Mistral ran a two-week launch discount during the preview, and the new price roughly matches that discounted rate. Anyone who paid list price will see the cut; early users who got the discount will barely notice it.

According to The Decoder, Artificial Analysis’s first independent test found the model fast and fairly cheap, but trailing Chinese open models overall. Mistral’s own comparison leaves out Kimi K3.1 Pro (weights due by 4 Nov), and Artificial Analysis still considers the per-token price on the high side for an open model of this size.

Where to Try It

The model is available on Mistral Studio (console.mistral.ai), the Mistral API, OpenRouter, Microsoft Foundry, NVIDIA-hosted endpoints, and Mistral’s coding agent Vibe at chat.mistral.ai/code. Hugging Face weights can be run with vLLM, SGLang, or llama.cpp GGUFs. Mistral does not state whether it is the default model in the Le Chat app.

The Arabic Angle

The Hugging Face model card lists Arabic (ar) and Persian (fa), and Mistral claims the model is natively fluent in 160+ languages. However, Mistral publishes no Arabic-specific results and nothing on Arabic dialects; readers should not assume any verified Arabic quality. Furthermore, API prices appear in USD (with an EUR documentation toggle), with no regional pricing mentioned for Arab nations.

To verify capabilities independently, run a quick test of your own before you rely on it:

• Give the model the same task in Modern Standard Arabic and a regional dialect such as Egyptian or Gulf.
• Ask for the response in the same dialect to test conversational fluency.
• Paste a long Arabic PDF contract to evaluate summary accuracy and right-to-left formatting.

What to Do Now

Enterprise buyers needing sovereign open weights should evaluate the model through the API or Mistral Studio before budgeting for private GPU clusters. Developers building agent workflows should test Vibe Code at chat.mistral.ai/code to measure token efficiency. Teams that work in Arabic should treat the “160+ languages” line as a promise, not a result, and run the dialect and document tests above first.