Agents-A1-4B-MLX-4bit

Agents-A1-4B · MLX 4-bit

Unofficial MLX conversion for Tippi.

This repository is an unofficial text-only MLX conversion of InternScience/Agents-A1-4B. It is not an official InternScience release. The source checkpoint and this conversion are distributed under Apache-2.0; retain the upstream notices and attribution when redistributing.

Conversion

• Upstream revision: 945c40a4aa6f534d434a353207b8d42ecf7a5293
• Converter: MLX-LM 0.31.3
• Quantization: affine 4-bit
• Group size: 64
• Average bits per weight: 4.503

hf download InternScience/Agents-A1-4B \
  --revision 945c40a4aa6f534d434a353207b8d42ecf7a5293 \
  --local-dir Agents-A1-4B-source

mlx_lm.convert \
  --hf-path Agents-A1-4B-source \
  --mlx-path Agents-A1-4B-4bit \
  --quantize \
  --q-bits 4 \
  --q-group-size 64

The artifact contains the Qwen 3.5 text model only: 924 tensors under language_model, with no vision tower or projector weights.

Tippi usage

Tippi loads this model locally through MLX Swift LM 3.31.3. It applies the checkpoint chat template with enable_thinking=false, uses temperature 0, and does not provide tools, web search, filesystem access, or cloud inference.

Local reference measurement

On the development M4 Pro Mac, a Traditional Chinese non-thinking probe generated 160 tokens at 85.7 tokens/s, prefilling 58 tokens at 350 tokens/s, with 2.63 GB peak MLX memory and 3.28 seconds process time. These are machine-specific research observations, not release guarantees.

Attribution

• Source model: https://huggingface.co/InternScience/Agents-A1-4B
• MLX-LM: https://github.com/ml-explore/mlx-lm
• Tippi integration: https://github.com/wcAmon/catcher-tippi