Mistral releases the open weights of Mistral Large 4

The Hugging Face model card describes a natively multimodal Mixture-of-Experts model with about 1 trillion parameters, 49 billion of them active per token, or 52 billion counting embeddings and output layers[2]. It uses 128 routed experts with 4 active per token plus one shared expert, takes text and image input, returns text, and has a 1 million token context window[2]. Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and that it is natively fluent in more than 160 languages, including every official EU language[2].

One set of weights covers instruction following, reasoning, coding and agentic work[2]. The released version adds three reasoning modes, low, medium and high, with high as the default and low replacing the old "None" mode[1].

Mistral's headline claim is agentic coding. On Terminal-Bench 4.0 it says the model improved by 13 percentage points to 41% compared with the preview, taking the lead among open-weight models while cutting output tokens by 30%[1]. The Next Web puts that just ahead of GLM-5.3 at 40%[6].

On the Artificial Analysis Coding Agent Index, Mistral says the model scores 48%, effectively tied with GLM-5.3, while using 47% fewer tokens per task on average[1]. It also reports a leading 19.7% on Cognition's private FrontierCode 2.0 benchmark, which judges whether a maintainer would merge the agent's change, and says blind human raters in Surge's code evaluation preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking it behind only Claude Opus 5.5[1]. These are Mistral's figures; a full technical report is "coming soon"[1].

The other half of the launch is defensive security. Mistral says the model scores 51% on the Artificial Analysis Cyber Index, which tests finding, reproducing and patching vulnerabilities, and refuses 96.6% of malicious cyber requests on average across the cyber prompts of JailbreakBench, StrongREJECT and AgentHarm[1]. It quotes Booz Allen, Cisco, Cloudflare, Rubrik and Zscaler in support[1].

Mistral is also candid about a risk. It says that while monitoring live traffic since the preview, it saw the model refuse black-box attacks against live targets, but "a potential for uplift to conduct real attacks", and that it is changing its guardrails and will keep updating them[1]. A separate variant tuned for penetration testing and red-teaming is going only to vetted enterprise partners such as Orange Cyberdefense, under contract and restricted to authorised systems[1].

The weights are on Hugging Face in two official variants, FP8 and NVFP4, with GGUF builds from Unsloth and the llama.cpp team and day-zero support in transformers, vLLM and llama.cpp[4]. The model card says the checkpoint ships with an embedded speculative decoding draft head, and that the NVFP4 version targets Blackwell GPUs[2]. The Next Web reports the quantised version runs on four Nvidia B200 or B300 GPUs[6].

For hosted use, Mistral offers it in Mistral Studio, through its API as mistral-large-4, on OpenRouter, in Vibe Code, as an NVIDIA NIM container and through Microsoft Foundry[1][2]. API pricing is $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens[1].

The model ships under the Mistral Ecosystem Open-Weight License, a modified MIT license[3]. It allows use, modification, fine-tuning, distribution and sale. The one condition is that a company operating a Model-as-a-Service business, defined as giving third parties meaningful control over the model's inputs, parameters or training data through an API or hosted endpoint, must sign a separate commercial license with Mistral before using it commercially[3]. Internal use, products that embed the model in specific features, and use through Mistral's own products are exempt[3].

SiliconANGLE frames Mistral Large 4 as a Western-developed option for companies that want to self-host, in a market crowded with very large Chinese coding models such as Zhipu's GLM family and Moonshot's Kimi family[7]. For teams running coding agents for hours at a time, the token-efficiency claim matters as much as the scores: Mistral says the model matches GLM-5.3 on the Coding Agent Index while using 47% fewer tokens per task, though actual running cost also depends on each provider's token prices and hardware[1]. The caveats are that the benchmarks are Mistral's own until the technical report and independent tests land, and that hosting providers selling access need Mistral's commercial license[3].