3 versions from Mistral AI, reviewed by the TopReviewed AI panel.
| Model | Score | Status | Released | Price in/out | Context | |
|---|---|---|---|---|---|---|
| Ministral 3 14B best small open model for on-device reasoning | 7.9 | GA | 2025-12-01 | $0.20 / $0.20 | 256K | Review → |
| Ministral 3 8B fast multilingual edge model with vision | 7.5 | GA | 2025-12-01 | $0.15 / $0.15 | 256K | Review → |
| Ministral 3 3B on-device AI with native vision | 7.1 | GA | 2025-12-01 | $0.10 / $0.10 | 256K | Review → |
Ministral 3 14B (release 25.12, shipped 2 December 2025) is the top of Mistral's edge tier and the strongest small open model in the lineup: a 14B dense transformer (13.5B language model + 0.4B vision encoder) under Apache 2.0, with a 256K context and native vision. Its reasoning variant posts a remarkable AIME 2025 85.0% — best-in-class for sub-20B — alongside GPQA Diamond 71.2%, MATH 90.4%, LiveCodeBench 64.6%, and MMLU 79.4%. Symmetric pricing at $0.20/$0.20. The buyer's sentence: the best cost-to-performance small open model for single-GPU and on-device reasoning, with native vision rare at this size. - Provider: Mistral AI (Paris, France) - Release: 2025-12-02, status GA - Context: 256,000 tokens; max output 16,384 - Modalities: text + image in, text out (native multimodal) - Knowledge cutoff: ~September 2025 - Headline price: $0.20 input / $0.20 output per 1M tokens (symmetric) - Architecture: 14B dense (13.5B LM + 0.4B vision); base/instruct/reasoning variants, all Apache 2.0
Full review →