Mistral has released Mistral Large 4, nicknamed 'Le Chonk', a multimodal, multilingual Mixture-of-Experts model that accepts text and images and generates text. The model has about 1 trillion total parameters, roughly 49 billion of which are active for each token, and supports a context window of 1 million tokens. Weights are published on Hugging Face in two official variants, FP8 and NVFP4, and community GGUF builds from Unsloth and the llama.cpp team are also available. The model is distributed under the Mistral Ecosystem Open-Weight License, which is derived from MIT but limits model-as-a-service offerings. Cybersecurity is the model's most prominent strength. Mistral reports that Large 4 scores 82% on the Artificial Analysis Cyber Index, a test in which a model must reproduce a real vulnerability in open-source software and then patch it. Mistral says several closed frontier models score near zero because they decline to attempt the task. The company positions the model mainly for defensive work, including malware analysis, vulnerability prioritisation and writing detection rules, although it was not explicitly trained for those tasks. Hugging Face argues that open weights matter here because security work often resembles an attack to provider safety filters, and API-dependent teams can lose access during an incident. The post also says the model still refuses harmful requests, with an average refusal rate on cyber prompts that it describes as higher than any other open-weight model it tested. Architecturally, the model combines multi-head latent attention with a DeepSeek-style sparse attention indexer that lets each query attend to up to 2,048 key positions. Its feed-forward layers route each token to four of 128 experts, plus one shared expert that always runs. Image input is handled by a 1.6 billion parameter Pixtral vision encoder. Hugging Face says the model is the fifth-best open-weights model on the Artificial Analysis Intelligence Index as of October 29, and reports 61.7% on DeepSWE v1.1 and a combined Coding Agent Index score of 49.8%. These figures come from the post and have not been independently verified here. For Australian businesses and professionals, the practical impact is limited. Running a model of this size requires substantial multi-GPU infrastructure, so most firms would access it through a hosted provider rather than in-house. The licence restriction on model-as-a-service offerings matters if you plan to resell access. Security teams that cannot send sensitive logs or code to overseas services may find self-hosted open weights useful for meeting Privacy Act obligations, including APP 8 on cross-border disclosure, though this does not remove those obligations. For most people, nothing changes, and it is worth testing only if you run a security function or a defensive engineering workload.