Ministral 3 3B is the smallest member of the Ministral 3 family, a Mistral lineup launched on December 2, 2025 alongside larger Mistral Large 3, Devstral Small 2, and Devstral 2 releases. It is distributed as an open-weight model under the Apache License 2.0, with Ollama listings reporting roughly 3.85 billion parameters in a mistral3 architecture, default Q4_K_M quantization, and a 3.0 GB on-disk footprint. Mistral markets the variant for edge and on-device deployment, describing it as capable of running on laptops or modest 4–8 GB RAM hardware without a GPU, while still supporting structured output and basic coding assistance for boilerplate generation in Python, JavaScript, HTML/CSS, and SQL.
Independent benchmarks place Ministral 3 3B at the small tier of the lineup, with Opper AI scoring it at an Intelligence index of 7, sitting below the 8B (Intelligence 9) and 14B (Intelligence 11) siblings and well under Mistral Large 3's flagship score of 16. Amazon Bedrock and Mistral's own endpoint both serve the model, and Artificial Analysis finds Bedrock delivering substantially higher output throughput at 453.7 tokens per second versus Mistral's 188.8 t/s, while Mistral's direct endpoint responds with lower time-to-first-token latency of 0.62 seconds compared to Bedrock's 0.92 seconds. That trade-off makes Bedrock appealing for throughput-sensitive batch or streaming workloads, and the Mistral endpoint better suited to interactive applications, giving developers a practical choice between speed and responsiveness when integrating a lightweight open-weight model.
