Agents-A1-4B-mlx-bf16

ToPo-ToPo/Agents-A1-4B-mlx-bf16

MLX bf16 conversion of InternScience/Agents-A1-4B for Apple Silicon (mlx-vlm).

Agents-A1-4B is a Qwen3_5ForConditionalGeneration (Qwen3.5-VL family) vision-language model. The text tower uses a hybrid of linear attention and full attention (full_attention_interval: 4) with interleaved MRoPE.

Provenance (self-converted from official weights)

• Source: InternScience/Agents-A1-4B (license: apache-2.0)
• Tool: mlx-vlm 0.6.6 — mlx_vlm.convert --hf-path InternScience/Agents-A1-4B --mlx-path . --dtype bfloat16
• Effective: 16 bits/weight (unquantized)
• Size on disk: 8.5 GB — peak memory 9.9 GB, ~70 tok/s generation (M-series, short prompt)

Effective bit width is higher than nominal because embeddings and parts of the vision tower are kept at higher precision (mlx default behaviour).

Validation

Checked after conversion: correctly described a synthetic test image (red square / blue circle, both shape and colour), and answered a text-only arithmetic prompt correctly. This is a smoke test, not a benchmark — no eval suite was run.

Usage

from mlx_vlm import load, generate
model, processor = load("ToPo-ToPo/Agents-A1-4B-mlx-bf16")

CLI:

mlx_vlm.generate --model ToPo-ToPo/Agents-A1-4B-mlx-bf16 \
  --image path/to/image.png --prompt "Describe this image." --max-tokens 200

Requirements

mlx-vlm >= 0.6.6. Earlier versions do not implement the qwen3_5 architecture and cannot load or convert this model.

License

Apache-2.0, inherited from the base model. Weights were converted/quantized to MLX format; no other modification to the weights.

Note on config files

tokenizer_config.json and processor_config.json were re-serialized by transformers 5.14.1 during conversion, so they differ textually from upstream: the chat template now lives in chat_template.jinja (byte-identical to upstream), special tokens moved into model_specific_special_tokens, and the image processor uses min_pixels/max_pixels instead of size/resample. The tokenizer was verified functionally identical to upstream (identical vocab, and identical encode/decode on a spot-check set). Users on transformers 4.x should prefer the upstream config files.