MiniCPM5-1B MLX DWQ 4-bit

MLX 4-bit quantization of openbmb/MiniCPM5-1B.

Quantization

• Runtime: MLX / mlx-lm
• Bits: 4
• Group size: 64
• Mode: affine
• Observed packed size: about 580 MB
• Conversion log: mlx_lm.convert reported 4.501 bits per weight

Smoke Test

Local Apple M2 smoke test:

• Prompt tokens: 17
• Prompt speed: 28.621 tok/s
• Generation speed: 133.897 tok/s
• Peak memory: 0.674 GB

Validation Status

Runtime loading and generation work, but this artifact is not yet release-approved on the strict multilingual/code/tool-use validation matrix. Known failures include Persian response quality, Arabic arithmetic response quality, and reasoning leakage on tool-planning prompts.

Usage

pip install -U mlx-lm
mlx_lm.generate --model Reza2kn/MiniCPM5-1B-MLX-DWQ-4bit --prompt "Hello" --max-tokens 64

Licensing and attribution

This model distribution is licensed under the Apache License, Version 2.0. See LICENSE. Existing third-party copyright, license, and attribution notices remain applicable.

• Upstream: openbmb/MiniCPM5-1B; declared license: apache-2.0.