Bijan Bowen introduces Mistral Large 4 as a preview model and separates reported benchmark results from his own hands-on tests. He explains that its eventual open-weight releaseAn open-weight model makes its trained parameter values available so others can run or adapt it under the release's licence. does not make a trillion-parameter model practical for ordinary local hardware, and evaluates the preview through Mistral's Vibe coding harnessAn AI agent harness is the software framework that packages a model with tools, instructions, context management, execution controls, and user interaction..
Bijan Bowen's browser operating-system test produces working utilities and a notably creative mechanism for installing custom apps. Skateboarding, pool-party and subway games demonstrate useful coding capabilityCode generation uses AI or another automated system to create source code from instructions, examples, schemas, or higher-level specifications. but also show inverted text, broken interactions and modest visual quality. An image-guided revisionImage-to-code uses an AI system to interpret a screenshot, mockup, or visual reference and produce a corresponding software interface implementation. initially fails because the harness is not actually providing native image access, illustrating how tooling can confound a model evaluationEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria..
Bijan Bowen finds further weaknesses in watch-site scale, apartment geometry and reference-image replication, while a multi-era city captures some period details despite spatial errors. His roughly $47 test bill and mixed outputs lead to a qualified assessment: a significant improvement over earlier Mistral models, not proof of superiority over competitors or a finished production system.
Watch on YouTube




