Matthew Berman examines Mistral Large 4's public preview, discussing its mixture-of-experts design, multimodal input and reported coding, agentic and cybersecurity benchmarks. He contrasts open-weight control and European infrastructure with the performance and convenience of hosted frontier models, noting that the weights were not yet released at the time of recording.
In a hands-on Rubik's Cube task, Matthew Berman reports configuration trouble, excessive reasoning output and a visualization that still fails after iteration. He distinguishes the model's capabilities from the quality of its coding harness and argues that open models need easier integration if they are to reach users without deployment expertise.
Watch on YouTube
