mlx-community/Ministral-3-14B-Base-2512-bf16

This is, Ministral 3 14B Base 2512 is a vision-language model: a text backbone paired with a vision encoder, supporting image understanding alongside text. This is the base pre-trained checkpoint — not instruction- or chat-tuned. For chat/instruction-following use cases, use the Instruct variant instead; this base checkpoint is intended for custom post-training/fine-tuning.

Community note. Structural check confirms the vision tower and multimodal projector were carried over intact (not dropped, which is a real failure mode for text-only conversion tools on vision-language models). Functional check (running generation) could not be completed on the maintainer's hardware -- this bf16 (unquantized) checkpoint needs ~26.6GB of GPU memory to generate, which exceeds the practical ceiling on a 32GB Apple Silicon machine (reproducibly hits a Metal out-of-memory error, not a one-off fluke). The weights themselves are an unmodified dtype/format conversion (no quantization math applied), so there's no reason to expect a correctness issue -- but generation hasn't actually been run end-to-end. If you have a Mac with more unified memory (64GB+), please open a discussion with your results.

This is an MLX conversion of mistralai/Ministral-3-14B-Base-2512, converted with mlx-vlm. Refer to the original model card for the full description, capabilities, and license terms.

Heads up

• Base model, not instruct-tuned — expect raw completion behavior, not chat-following. Don't expect it to follow instructions well.
• Vision retained at full precision — only the language backbone is quantized; the vision tower and multimodal projector are untouched bf16, per mlx-vlm's standard policy of not quantizing multimodal modules.
• Output size on disk: 27.92GB

Provenance

• Source: mistralai/Ministral-3-14B-Base-2512 (BF16)
• Language model layers: 16 (bf16, unquantized)-bit affine quantization, group_size=n/a
• Vision tower + multimodal projector: kept at full precision (not quantized)
• Blended average: 16 (bf16, unquantized) bits per weight across all parameters

Ministral 3 family

Model	Type	mlx-community (4-bit)
Ministral 3 3B Base 2512	Base pre-trained	mlx-community/Ministral-3-3B-Base-2512-4bit
Ministral 3 3B Instruct 2512	Instruct post-trained	mlx-community/Ministral-3-3B-Instruct-2512-4bit
Ministral 3 3B Reasoning 2512	Reasoning capable	mlx-community/Ministral-3-3B-Reasoning-2512-4bit
Ministral 3 8B Base 2512	Base pre-trained	mlx-community/Ministral-3-8B-Base-2512-4bit
Ministral 3 8B Instruct 2512	Instruct post-trained	mlx-community/Ministral-3-8B-Instruct-2512-4bit
Ministral 3 8B Reasoning 2512	Reasoning capable	mlx-community/Ministral-3-8B-Reasoning-2512-4bit
Ministral 3 14B Base 2512	Base pre-trained	mlx-community/Ministral-3-14B-Base-2512-4bit
Ministral 3 14B Instruct 2512	Instruct post-trained	mlx-community/Ministral-3-14B-Instruct-2512-4bit
Ministral 3 14B Reasoning 2512	Reasoning capable	mlx-community/Ministral-3-14B-Reasoning-2512-4bit

Use with mlx

pip install -U mlx-vlm

python -m mlx_vlm.generate --model mlx-community/Ministral-3-14B-Base-2512-bf16 --max-tokens 100 --temperature 0.0 --prompt "Describe this image." --image <path_to_image>

For text-only prompts, omit --image.

License

Apache 2.0 — see the original model card for the full license text and any usage terms.