Open-Source Model Stack 2026

This paper compares named open-weight model snapshots: Llama 4 Scout (109B total parameters, 17B active), Qwen3 235B-A22B, Mistral Small 4 (119B total), and DeepSeek V3-0324. Total and active parameter counts describe different things; deployment memory, latency, and license requirements must be evaluated for the selected model and setup. Meta model announcement; Mistral model documentation.

The problem is that choosing between them requires navigating a complex space of license terms, inference cost profiles, fine-tuning behavior, language coverage, and compliance implications. A model that is optimal for a Taiwanese financial institution's document processing workflow is not the same model that is optimal for a Japanese hospital's clinical summarization use case.

This whitepaper presents a model-selection framework covering license terms, inference costs, language coverage, and deployment constraints. Test candidate models against representative tasks and consult the current model cards before choosing a deployment.