Mistral has published updated open weights for Mistral Large 4, its flagship model, positioned for agentic coding and cybersecurity work and customisable through Mistral Forge. According to Artificial Analysis, the model has 1.05 trillion total parameters, 52 billion active, a 1 million token context window and accepts text and image input. Mistral is pricing it at $0.70 per million input tokens and $2.10 per million output tokens, about half the price of the preview. Its MEOW licence requires businesses that offer the model as a hosted service to obtain a separate commercial licence. Independent benchmarks put the model in a middle tier. Artificial Analysis scores it 26 on its Intelligence Index v5, making it the strongest open-weight model developed outside China, but it trails DeepSeek V4.1 Flash 1017 and GLM-5.3 by six points. On Cognition's FrontierCode 2.0 it scores 15%, compared with 29% for DeepSeek and 22% for GLM-5.3, and it reaches 25% on Terminal-Bench 4.0. Mistral says it uses 39% fewer output tokens per task than the preview, and its hallucination rate fell from 41% to 28%, though accuracy remains low at 26%. Mistral's own claim of a 64% human-preference win rate against GLM and Kimi models has not been independently verified. Serving support arrived on day zero through vLLM, and Unsloth has released a 1-bit quantised version of about 223GB, compared with 1.1TB for the full FP8 checkpoint. For Australian businesses and professionals, the release changes little immediately. The lower API price makes it a cheaper option to test, but the full-precision model is too large for most local infrastructure, so quantised versions are the realistic route for experimentation. Any firm planning to resell the model as a service should check the MEOW licence terms first. If customer or personal information is sent to a hosted API, obligations under the Privacy Act 1988, including the rules on cross-border disclosure, apply, and the provider's data location should be confirmed. It is worth trialling on non-sensitive coding tasks, but the benchmarks do not support treating it as the best available option.