Mistral: Ministral 3 8B 2512 (batch)

Ministral 3 8B 2512 (batch) is Mistral AI’s batched API offering for the Ministral 3 8B instruct checkpoint released in December 2025. It is a dense, Apache 2.0 open-weight multimodal model built via cascade distillation from Mistral Small 3.1, combining an ~8.4B language backbone with a ~0.4B vision encoder for text-and-image understanding. The family targets efficient edge and local deployment while supporting a very long context window, native tool calling, structured JSON output, and broad multilingual coverage.

The batch endpoint exposes the same capabilities as the standard Ministral 8B instruct API—vision, agentic tool use, and chat/instruction following—at reduced per-token cost for asynchronous workloads. Weights are published on Hugging Face in FP8 and BF16 variants (alongside separate Base and Reasoning checkpoints), and the model is designed to run on a single consumer or datacenter GPU when quantized appropriately.

On public evaluations, Ministral 3 8B delivers strong results for its size class on knowledge and math benchmarks, competitive multimodal scores, and solid code metrics on the dedicated Reasoning variant, while API-side leaderboards (e.g., Artificial Analysis) report instruct-style scores on MMLU-Pro, GPQA, HLE, and agentic suites. It is well suited to local assistants, document and image QA, translation, and lightweight agent pipelines where cost and latency matter.